![Banner image](https://cdn-prod.scdn6.secure.raxcdn.com/static/media/859ff7c0-8004-4274-81cb-3e20fe725291.jpeg?_cb=1710969314.762538)
Lab Insight: Dell AI PoC for Transportation & Logistics
Read the Report Executive Summary View the InfographicWed, 20 Mar 2024 21:23:12 -0000
|Read Time: 0 minutes
Introduction
As part of Dell’s ongoing efforts to help make industry-leading AI workflows available to its clients, this paper outlines a sample AI solution for the transportation and logistics market. The reference solution outlined in this paper specifically targets challenges in the maritime industry by creating an AI powered cargo monitoring PoC built with DellTM hardware.
AI as a technology is currently in a rapid state of advancement. While the area of AI has been around for decades, recent breakthroughs in generative AI and large language models (LLMs) have led to significant interest across almost all industry verticals, including transportation and logistics. Futurum intelligence projects a 24% growth of AI in the transportation industry in 2024 and a 30% growth for logistics.
The advancements in AI now open significant possibilities for new value-adding applications and optimizations, however different industries will require different hardware and software capabilities to overcome industry specific challenges. When considering AI applications for transportation and logistics, a key challenge is operating at the edge. AI-powered applications for transportation will typically be heavily driven by on-board sensor data with locally deployed hardware. This presents a specific challenge, requiring hardware that is compact enough for edge deployments, powerful enough to run AI workloads, and robust enough to endure varying edge conditions.
This paper outlines a PoC for an AI-based transportation and logistics solution that is specifically targeted at maritime use cases. Maritime environments represent some of the most rigorous edge environments, while also presenting an industry with significant opportunity for AI-powered innovation. The PoC outlined in this paper addresses the unique challenges of maritime focused AI solutions with hardware from Dell and BroadcomTM.
The PoC detailed in this paper serves as a reference solution that can be leveraged for additional maritime, transportation, or logistics applications. The overall applicability of AI in these markets is much broader than the single maritime cargo monitoring solution, however, the PoC demonstrates the ability to quickly deploy valuable edge-based solutions for transportation and logistics using readily available edge hardware.
Importance for the Transportation and Logistics Market
Transportation and logistics cover a broad industry with opportunity for AI technology to create a significant impact. While the overarching segment is widespread, including public transportation, cargo shipping, and end-to-end supply chain management, key to any transportation or logistics process is optimization. These processes are dependent on a high number of specific details and variables such as specific routes, number and types of goods transported, compliance regulations, and weather conditions. By optimizing for the many variables that may arise in a logistical process, organizations can be more efficient, save money, and avoid risk.
In order to create these optimal processes, however, the data surrounding the many variables involved needs to be captured. Further, this data needs to be analyzed, understood, and acted on. The large quantity of data required and the speed at which it must be processed in order to make impactful decisions to complex logistical challenges often surpasses what a human can achieve manually.
By leveraging AI technology, impactful decisions to transportation and logistics processes can be achieved quicker and with greater accuracy. Cameras and other sensors can capture relevant data that is then processed and understood by an AI model. AI can quickly process vast amounts of data and lead to optimized logistics conclusions that would otherwise be too timely, costly, or complex for organizations to make.
The potential applications for AI in transportation are vast and can be applied to various means of transportation including shipping, rail, air, and automotive, as well as associated logistical processes such as warehouses and shipping yards. One possible example is AI optimized route planning which could pertain to either transportation of cargo or public transportation and could optimize for several factors including cost, weather conditions, traffic, or environmental impact. Additional applications could include automated fleet management, AI powered predictive vehicle maintenance, and optimized pricing. As AI technology improves, many transportation services may be additionally optimized with the use of autonomous vehicles.
By adopting such AI-powered applications, organizations can implement optimizations that may not otherwise be achievable. While new AI applications show promise of significant value, many organizations may find adopting the technology a challenge due to unfamiliarity with the new and rapidly advancing technology. Deploying complex applications such as AI in transportation environments can pose an additional challenge due to the requirements of operating in edge environments.
The following PoC solution outlines an example of a transportation focused AI application that can offer significant value to maritime shipping by providing AI-powered cargo monitoring using Dell hardware at the edge.
Solution Overview
To demonstrate an AI-powered application focused on transportation and logistics, Scalers AITM, in partnership with Dell, Broadcom, and The Futurum Group implemented a proof-of-concept for a maritime cargo monitoring solution. The solution was designed to capture sensor data from cargo ships as well as image data from on-board cameras. Cargo containers can be monitored for temperature and humidity to ensure optimal conditions are maintained for the shipped cargo. In addition, cameras can be used to monitor workers in the cargo area to ensure worker safety and prevent injury. The captured data is then utilized by an LLM to create an AI generated compliance report at the end of the ship’s voyage.
This proof-of-concept addresses several problems that can be encountered in maritime shipping. Refrigerated cargo, known as reefer, is utilized to ship perishable items and pharmaceuticals that must be kept at specific temperatures. Without proper monitoring to ensure optimal temperatures, reefer may experience swings in temperature, resulting in spoiled products and ultimately financial loss. Predictive forecasting of the power requirements for refrigerated cargo can provide additional cost and environmental savings by providing greater power usage insights.
Similarly, dry cargo can become spoiled or damaged when exposed to excessive moisture. Moisture can be introduced in the form of condensation – known as cargo sweat – due to changes in climate and humidity during the ships journey. By monitoring the temperature and humidity of the cargo, alerts can be raised signaling the possibility of cargo sweat and allowing ventilation adjustments to be made which can prevent moisture related damage.
A third issue addressed by the maritime cargo monitoring PoC is that of worker safety. The possibility of shifting cargo containers can lead to dangerous situations and potential injuries for those working in container storage areas. By using video surveillance of workers in cargo areas, these potential injuries can be avoided.
The PoC provides monitoring of these challenges with an additional visualization dashboard that displays information such as number of cargo containers, forecasted energy consumption, container temperature and humidity, and a video feed of workers. The dashboard additionally raises alerts as issues arise in any of these areas. This information is further compiled in to an end of voyage report for compliance and logging purposes, automatically generated with an LLM.
To achieve the PoC solution, simulated sensor data is generated for both reefers and dry containers, approximating the conditions undergone during a real voyage. The sensor data is written to an OPCUA server which then supplies data to a container sweat analytics module and a power consumption predictor. For dry containers, the temperature and humidity data is utilized alongside the forecasted weather of the route to create dew point calculations and monitor potential container sweat. Sensor data recording the temperature of reefer containers is monitored to ensure accurate temperatures are maintained, and a decision tree regressor model is leveraged to predict future power consumption for the next hour.
Figure 1: Visualization Dashboard
For monitoring worker safety, RTSP video data is captured into a video analytics pipeline built on NVIDIATM DeepStream. Streaming data is decoded and then inferenced using the YoloV8s model to detect workers entering dangerous, restricted zones. The restricted zones are configured as x,y coordinate pairs stored as JSON objects. Uncompressed video is then published to the visualization service using the Zero Overhead Network Protocol (Zenoh).
Monitoring and alerts for all of these challenges is displayed on a visualization dashboard as can be seen in Figure 1, as well as summarized in an end of voyage compliance report. The resulting compliance report that details the information collected on the voyage is AI generated using the Zephyr 7B model. Testing of the PoC found that the report could be generated in approximately 46 seconds, dramatically accelerating the reporting process compared to a manual approach.
To achieve the PoC solution in-line with the restraints of a typical maritime use case, the solution was deployed using Dell PowerEdge servers designed for the edge. The sensor data calculations and predictions, video pipeline, and AI report generation were achieved on a Dell PowerEdge XR7620 server with dual NVIDIA A100 GPUs. A Dell PowerEdge XR12 server was deployed to host the visualization dashboard. The two servers were connected with high bandwidth Broadcom NICs.
An overview of the solution can be seen in Figure 2
Additional details about the implementation and performance testing of the PoC on GitHub, including:
- Configuration information including diagrams and YAML code
- Instructions for doing the performance tests
- Details of performance results
- Source code
- Samples for test process
https://github.com/dell-examples/generative-ai/tree/main/transportation-maritime
Highlights for AI Practitioners
The cargo monitoring PoC demonstrates a solution that can avoid product loss, enhance compliance and logging, and improve worker safety, all by using AI. The creation of these AI processes was done using readily available AI tools. The process of creating valuable, real-world solutions by utilizing such tools should be noted by AI practitioners.
The end of voyage compliance report is generated using the Zephyr 7B LLM model created by Hugging Face’s H4 team. The Zephyr 7B model, which is a modified version of Mistral 7B, was chosen as it is a publicly available model that is both lightweight and highly accurate. The Zephyr 7B model was created using a process called Distilled Supervised Fine Tuning (DSFT) which allows the model to provide similar performance to much larger models, while utilizing far fewer parameters. Zephyr 7B, which is a 7 Billion parameter model, has demonstrated performance comparable to 70 Billion parameter models. This ability to provide the capabilities of larger models in a smaller, distilled model makes Zephyr 7B an ideal choice for edge-based deployments with limited resources, such as in maritime or other transportation environments.
While Zephyr 7B is a very powerful and accurate LLM model, it was trained on a broad data set and it is intended for general purpose usage, rather than specific tasks such as generating a maritime voyage compliance report. In order to generate a report that is accurate to the maritime industry and the specific voyage, more context must be supplied to the model. This was achieved using a process called Retrieval Augmented Generation (RAG). By utilizing RAG, the Zephyr 7B model is able to incorporate the voyage specific information to generate an accurate report which detailed the recorded container and worker safety alerts. This is notable for AI practitioners as it demonstrates the ability to use a broad, pre-trained LLM model, which is freely available, to achieve an industry specific task.
To provide the voyage specific context to the LLM generated report, time series data of recorded events, such as container sweating, power measurements, and worker safety violations, is queried from InfluxDB at the end of the voyage. This text data is then embedded using the Hugging Face LangChain API with the gte-large embedding model and stored in a ChromaDB vector database. These vector embeddings are then used in the RAG process to provide the Zephyr 7B model with voyage specific context when generating the report.
AI practitioners should also note that AI image detection is utilized to detect workers entering into restricted zones. This image detection capability was built using the YOLOv8s object detection model and NVIDIA DeepStream. YOLOv8s is a state of the art, open source, AI model for object detection built by Ultralytics. The model is used to detect workers within a video frame and detect if they enter into pre-configured restricted zones. NVIDIA DeepStream is a software development toolkit provided by NVIDIA to build and accelerate AI solutions from streaming video data, which is optimized for NVIDIA hardware such asthe A100 GPUs used in this PoC. It is notable that NVIDIA DeepStream can be utilized for free to build powerful video-based AI applications, such as the worker detection component of the maritime cargo monitoring solution. In this case, the YOLOv8s model and the DeepStream toolkit are utilized to build a solution that has the potential to prevent serious workplace injuries.
Key Highlights for AI Practitioners
- Maritime compliance report generated with Zephyr 7B LLM model
- Retrieval Augmented Generation (RAG) approach used to provide Zephyr 7B with voyage specific information
- YOLOv8s and NVIDIA DeepStream used to create powerful AI worker detection solution using video streaming data
Considerations for IT Operations
The maritime cargo monitoring PoC is notable for IT operations as it demonstrates the ability to deploy a powerful AI driven solution at the edge. For many in IT, AI deployments in any setting may be a challenge, due to overall unfamiliarity with AI and its hardware requirements. Deployments at the edge introduce even further complexity.
Hardware deployed at the edge requires additional considerations, including limited space and exposure to harsh conditions, such as extreme temperature changes. For AI applications deployed at the edge, these requirements must be maintained, while simultaneously providing a system powerful enough to handle such a computationally intensive workload.
For the maritime cargo monitoring PoC, Dell PowerEdge XR7620 and PowerEdge XR12 servers were chosen for their ability to meet both the most demanding edge requirements, as well as the most demanding computational requirements. Both servers are ruggedized and are capable of operating in temperatures ranging from -5°C to 55°C, as well as withstanding dusty or otherwise hazardous environments. They additionally offer a compact design that is capable of fitting into tight environments. This provides servers that are ideal for a demanding edge environment, such as in maritime shipping, which may experience large temperature swings and may have limited space for servers. Meanwhile, the Dell PowerEdge XR7620 is also equipped with NVIDIA GPUs, providing it with the compute power necessary to handle AI workloads.
Dell PowerEdge XR7620
NVIDIA A100 GPUs were chosen as they are well suited for various types of AI workflows. The PoC includes both a video classification component and a large language model component, requiring hardware that is well suited for both workloads. While there are other processors that are more specialized specifically for either video processing or language models, the A100 GPU provides flexibility to perform both well on a single platform.
The use of high bandwidth Broadcom NICs is also a notable component of the PoC solution for IT operations to be aware of. The Broadcom NICs are responsible for providing a high bandwidth Ethernet connection between the cargo and worker monitoring applications and the visualization and alerting dashboard. The use of scalable, high bandwidth NICs is crucial to such a solution that requires transmitting large amounts of sensor and video data, which may include time sensitive information.
Detection of issues with either reefer or dry containers may require quick action to protect the cargo, and quick detection of workers in hazardous environments can prevent serious harm or injury. The use of a high bandwidth Ethernet connection ensures that data can be quickly transmitted and received by the visualization dashboard for operators to respond to alerts as they arise.
Key Highlights for IT Operations
- AI solution deployed on rugged Dell PowerEdge XR7620 and PowerEdge XR12 servers to accommodate edge environmentand maintain high computational requirements.
- NVIDIA A100 GPUs provide flexibility to support both video and LLM workloads.
- Broadcom NICs provide high bandwidth connection between monitoring applications and visualization dashboard.
Solution Performance Observations
Key to the performance of the maritime cargo monitoring PoC is its ability to scale to support multiple concurrent video streams for monitoring worker safety. The solution must be able to quickly decode and inference incoming video data to detect workers in restricted areas. The ability for the visualization dashboard to quickly receive this data is additionally critical for actions to be taken on alerts as they are raised. The solution was separated into a distinct inference server, to capture and inference data, and an encode server, to display the visualization service. This architecture allows the solution to scale the services independently as needed for varying requirements of video streams and application logic. The separate services are then connected with high bandwidth Ethernet using Broadcom NetXtreme®-E Series Ethernet controllers. The following performance data demonstrates the ability to scale the solution with an increasing number of data streams. Each test was run for a total of 10 minutes and video streams were scaled evenly across the two NVIDIA A100 GPUs. Additional performance results are available in the appendix.
Figure 4: Transportation PoC Throughput
Figure 4 displays the total throughput of frames per second as well as the average throughput as the number of streams increased. The frames per second metric includes video decoding, inference, post-processing, and publishing of an uncompressed stream. The PoC displayed increasing throughput with a maximum of 653.7 frames per second when tested with 24 concurrent streams. Notably, the average frames per second remained steady at approximately 30 frames per second for up to 20 streams, which is considered an industry standard for video processing workloads. When tested with 24 streams, the solution did experience a slight drop, with an average of 27.24 frames per second. Overall, the throughput performance demonstrates the ability of the Dell PowerEdge Server and the NVIDIA A100 GPUs to successfully handle a demanding video-based AI workload with a significant number of concurrent streams.
Figure 5: Transportation PoC Bandwidth Utilization
Figure 5 displays the solution’s bandwidth utilization as the number of streams increased from 1 to 24. The results demonstrate the increase in required bandwidth, both at a maximum and on average, as the number of streams increased. The average bandwidth utilization scaled from 1.56 Gb/s with a single video stream, to 34.6 Gb/s when supporting 24 concurrent streams The maximum bandwidth utilization was observed to be 3.13 Gb/s with a single stream, up to 67.9 Gb/s with 24 streams. By utilizing scalable, high bandwidth 100Gb/s Broadcom Ethernet, the solution is able to achieve the increasing bandwidth utilization required when adding additional video streams.
The performance results showcase the PoC as a flexible solution that can be scaled to accommodate varying levels of video requirements while maintaining performance and scaling bandwidth as needed. The solution also provides the foundation for additional AI-powered transportation and logistics applications that may require similar transmission of sensor and video data.
Final Thoughts
The maritime cargo monitoring PoC provides a concrete example of how AI can improve transportation and logistics processes by monitoring container conditions, detecting dangerous working environments, and generating automated compliance reports. While the PoC presented in this paper is limited in scope and executed using simulated sensor datasets, the solution serves as a starting point for expanding such a solution and a reference for developing related AI applications.
The solution additionally demonstrates several notable results. The solution utilizes readily available AI tools including Zephyr 7B, YOLOv8s, and NVIDIA DeepStream to create valuable AI applications that can be deployed to provide tangible value in industry specific environments. The use of RAG in the Zephyr 7B implementation is especially notable, as it provides customization to a general-purpose language model, enabling it to function for a maritime specific use case. The PoC also showcased the ability to deploy an AI solution in demanding edge environments with the use of Dell PowerEdge XR7620 and XR12 servers and to provide high bandwidth when transmitting critical data by using Broadcom NICs.
When tested, the PoC solution demonstrated the ability to scale up to 24 concurrent streams while experience little loss of throughput and successfully supporting increased bandwidth requirements. Testing of the LLM report generation showed that an AI augmented maritime compliance report could be generated in as little as 46 seconds. The testing of the PoC demonstrate both its real-world applicability in solving maritime challenges, as well as its flexibility to scale to individual deployment requirements.
Transportation and logistics are areas that rely heavily upon optimization. With the advancements in AI technology, these markets are well positioned to benefit from AI-driven innovation. AI is capable of processing data and deriving solutions to optimize transportation and logistics processes at a scale and speed that humans are not capable of achieving manually. The opportunity for AI to create innovative solutions in this market is broad and extends well beyond the maritime PoC detailed in this paper. By understanding the approach to creating an AI application and the hardware components used, however, organizations in the transportation and logistics market can apply similar solutions to innovate and optimize their business.
Appendix
Figure 6 shows full performance testing results for the cargo monitoring PoC.
CONTRIBUTORS
Mitch Lewis
Research Analyst | The Futurum Group
PUBLISHER
Daniel Newman
CEO | The Futurum Group
INQUIRIES
Contact us if you would like to discuss this report and The Futurum Group will respond promptly.
CITATIONS
This paper can be cited by accredited press and analysts, but must be cited in-context, displaying author’s name, author’s title, and “The Futurum Group.” Non-press and non-analysts must receive prior written permission by The Futurum Group for any citations.
LICENSING
This document, including any supporting materials, is owned by The Futurum Group. This publication may not be
reproduced, distributed, or shared in any form without the prior written permission of The Futurum Group.
DISCLOSURES
The Futurum Group provides research, analysis, advising, and consulting to many high-tech companies, including those mentioned in this paper. No employees at the firm hold any equity positions with any companies cited in this document.
ABOUT THE FUTURUM GROUP
The Futurum Group is an independent research, analysis, and advisory firm, focused on digital innovation and market-disrupting technologies and trends. Every day our analysts, researchers, and advisors help business leaders from around the world anticipate tectonic shifts in their industries and leverage disruptive innovation to either gain or maintain a competitive advantage in their markets.
© 2024 The Futurum Group. All rights reserved.
Related Documents
![Post thumbnail](https://cdn-prod.scdn6.secure.raxcdn.com/static/media/8b07fb53-7803-42c9-af8c-02d40a3ca11c.png?_cb=1715631463.9694412)
Lab Insight: Dell CPU-Based AI PoC for Retail
Mon, 13 May 2024 20:45:53 -0000
|Read Time: 0 minutes
Introduction
As part of Dell’s ongoing efforts to help make industry-leading AI workflows available to its clients, this paper outlines a sample AI solution for the retail market. The PoC leverages DellTM technology to showcase an AI-powered inventory management application for retail organizations.
AI technology has been in development for some time, but recent technological advancements have greatly accelerated AI’s ability to provide value across a wide range of enterprise applications. AI solutions have become a key initiative for many organizations. While the advancement of AI technology provides the basis for a diverse set of AI-powered applications, the specific requirements of different verticals provide distinct hardware and software challenges. IT organizations might be unsure of the technical requirements for deploying such a solution. This uncertainty may be due to unfamiliarity with AI, as well as an expectation that AI applications will require specialized hardware, often with limited availability.
This paper covers a solution specifically designed to capture the requirements of a retail-based AI deployment using a standard AMDTM CPU for AI training and inference. The solution leverages hardware from Dell, AMD , and BroadcomTM, to create a solution powerful enough to capture and analyze large-scale video data from cameras in retail environments, as well as flexible enough to scale to the unique needs of individual retail environments. Training of the model was achieved in two days, utilizing the same Dell PowerEdge server that is used for inferencing. The scalability of the solution was tested with up to 20 video streams. The PoC additionally demonstrates AI optimizations for AMD CPUs by utilizing AMD’s ZenDNN library. The utilization of the ZenDNN library, along with node pinning, resulted in an average throughput increase of 1.5x.
While the overall applications of AI in retail environments are much broader than the single inventory management solution outlined in this paper, the PoC demonstrates a framework for how IT organizations can quickly deploy an AI solution that delivers practical value in a retail environment by using readily available hardware.
Importance for the Retail Market
As with many other industries, the retail market has become increasingly data driven. Data can provide greater insight into areas such as customer behavior and product demand, as well as assist in optimizing operational areas such as procurement and inventory management. The emergence of AI technology provides even greater opportunity for valuable data-driven insights and optimizations within the retail industry.
Possibilities for retail-focused AI solutions include both customer experience (CX)-driven solutions and operations-focused applications. CX might be enhanced with personalized recommendation systems based on customer purchase trends, or virtual assistants capable of providing product recommendations for online retail experiences. Retail operations may be optimized through solutions such as AI-enhanced surveillance to detect fraud or theft, inventory management systems, or AI-powered product pricing systems.
These examples, as well as the more in-depth PoC study outlined in this paper, are a small subset of possible AI applications that may be implemented by retail organizations. While the exact solution implementations that are most appropriate may vary between organizations based on several factors such as location, size, type of goods sold, and distribution of online versus in-person sales, it is clear that AI applications can provide immense value in retail environments.
While a proactive approach to AI adoption may be beneficial to retail organizations, unfamiliarity with AI technology and the hardware and software components needed to deploy and optimize such solutions act as a barrier to adoption. The following solution demonstrates a PoC for an AI-powered retail inventory management system that can be quickly deployed and further expanded upon by retail organizations using commonly available hardware.
Solution Overview
The retail inventory management solution addresses a common challenge in retail environments of inventory distortion. Without accurate and timely inventory management, retail organizations can be challenged with stock levels that are either too low or too high. Both situations can prove to be costly. Too much inventory requires additional storage, commitment of capital, and potential waste of perishable items. Conversely, too low of inventory can lead to customer dissatisfaction and loss of sales. In many cases, low inventory leads to customers purchasing at competitive retailers and may lead to overall loss of brand loyalty. By utilizing computer vision and object detection AI models to monitor and track inventory, retailers can achieve real-time insights into their stock to balance their inventory more appropriately and provide valuable insights back to suppliers.
To demonstrate a real-world example solution of an AI application that could be deployed to address such retail challenges, Scalers AITM, in partnership with Dell, Broadcom, and The Futurum Group, implemented a PoC solution for a retail inventory management system. The solution was designed to capture data from store cameras and use an object-detection AI model to monitor and manage product stock levels. The solution was capable of detecting products on store shelves, keeping track of inventory, and raising alerts of low or out of stock items.
All of this was accomplished using standard Dell PowerEdge servers with 32 core 4th Gen AMD EPYC processors and Broadcom networking. No GPUs were required. The CPU-based solution was further optimized with AMD’s Zen Deep Neural Network (ZenDNN) library, which provides optimizations for deep learning inferencing on AMD CPU hardware. AMD’s ZenDNN optimizations delivered an average of 1.5x increased throughput performance to the PoC. By utilizing modest, CPU-based hardware, this PoC solution demonstrates a clear example of a readily deployable and broadly applicable AI retail solution.
To achieve the solution, store shelves were configured in zones with the product names and corresponding x,y coordinate pairs that indicated the shelf location. The products, location, and the maximum capacity for each item were stored as JSON objects.
Solution Highlights
|
Figure 1: Visualization Dashboard.
The identification and monitoring of products in each zone is achieved by capturing video data from store cameras into a video pipeline for processing. The live video stream is captured, decoded, and then inferenced using an object-detection AI model. The video pipeline is run on a typical Dell PowerEdge server without requiring any GPUs or specialized accelerators. The video streams can additionally be directed to Dell PowerScale NAS storage for long term retention. Zenoh (Zero Overhead Network Protocol) is then utilized for distribution to an additional Dell server running a visualization process. The visualization engine enables the video stream to be shared over the web for remote viewing and analysis. The visualization dashboard can be seen in Figure 1. Figure 2 depicts a high-level diagram of the solution pipeline.
Figure 2: Retail Inventory Management AI Pipeline (Source: Scalers AI)
By separating the architecture into two distinct pieces, with one server powering video decoding and object detection, and a separate server for the visualization process, the PoC provides a framework for a highly scalable solution. Traditional approaches would combine the processes into a single pipeline, however, this architecture can prove challenging to scale due to the different computational needs of the services. Utilizing a dual service approach, provides greater flexibility to scale the processes as needed for retail organizations further expanding upon this PoC. Both the video pipeline and the visualization service can be scaled independently as requirements such as the number of video streams or application logic are adjusted. The dual service architecture and scalability of the overall solution is enabled by utilizing high speed Broadcom NetXtreme-E NICs which maintain high bandwidth between the video inferencing and visualization services.
Additional details about the implementation and performance testing of the PoC have been made available by Dell on GitHub.
The key hardware components used in the solution include the following: Dell PowerEdge R7615 Servers
|
Highlights for AI Practitioners
It is notable for AI practitioners that the project was not limited to the deployment and inferencing of the AI model. The solution additionally involved customization of the pre-trained base model using a process known as Transfer Learning. The solution began with the SSD_MobileNet_v2 model for object detection, which was an ideal model for this PoC as it provides a one-stage object detection model that does not require exceptional compute power. The model was then customized via Transfer Learning with the SKU110K image data set. The training process involved 23,000 images and resulted in a mean average precision (mAP) of 0.7. The training process was completed in approximately two days.
Figure 3: Object Detection Software Overview
It should also be noted that both the model training and deployment of the video pipeline solution were accomplished using the same 32 core Dell PowerEdge R7615 server. The PoC demonstrates the ability to achieve useful AI applications on CPU-based hardware that is commonly found in retail environments. The solution is further optimized for inferencing on AMD CPUs by utilizing AMD’s ZenDNN library and node pinning. The ZenDNN library provides performance tuning for deep learning inferencing on AMD CPUs while node pinning can further optimize the application by binding processes to dedicated compute resources.
The below table shows the ZenDNN parameter configurations used.
Variable | Value | Notes |
TF_ENABLE_ZENDNN_OPTS | 0 | Sets native TensorFlow code path |
ZENDNN_CONV_ALGO | 3 | Direct convolution algorithm with blocked inputs and filters |
ZENDNN_TF_CONV_ADD_FUSION_SAFE | 0 | Default Value |
ZENDNN_TENSOR_POOL_LIMIT | 512 | Set to 512 to optimize for Convolutional Neural Network |
OMP_NUM_THREADS | 32 | Sets threads to 32 to match # of cores |
GOMP_CPU_AFFINITY | 0-31 | Binds threads to physical CPUs. Set to number of cores in the system |
Figure 4: ZenDNN Configurations
Key Highlights for AI Practitioners
|
Considerations for IT Operations
The hardware used in this AI application, including Dell PowerEdge R7615 servers with 4th Gen 32 core AMD EPYC 9354P Processors, Dell PowerScale NAS, Dell PowerSwitch Z9664, and Broadcom NetXtreme-E NICs, is familiar and available to IT operations, yet each component provides valuable characteristics needed to support this type of solution.
The Dell PowerEdge servers provide powerful 4th Generation AMD EPYC processors that are capable of supporting both the AI and application workloads, and the Dell PowerScale NAS provides a high-performance, highly scalable NAS storage system capable of handling large-scale video and image data. The solution is then tied together using Broadcom Ethernet capable of supporting the high bandwidth requirements of video streaming. Most notably, these components all provide scalability for IT organizations to further build out this application with more demanding requirements such as additional video streams or additional application logic.
Futurum Group Comment: The specific use of Dell PowerEdge R7615 servers should be noted, as it demonstrates the ability to run AI workloads on standard hardware, commonly deployed in retail environments. While not considered a high-end compute server, the R7615 servers with mid-range 32 core 9354P Processors proved capable of all processes including model training, inferencing, and the separate visualization engine. This enables retail IT organizations to deploy such solutions without acquiring GPUs or requiring the datacenter level cooling needed for higher end servers. Additionally, by separating the architecture into separate video and visualization pipelines, the solution can be scaled to meet the size and performance requirements of a broad range of retail environments.
The on-premises deployment of this solution additionally enables IT operations to achieve their data security and data privacy requirements. While public cloud has been utilized for many early iterations of AI applications, data privacy becomes a concern for many organizations as they build further AI applications leveraging private data. By deploying this, or similar, retail solutions on-premises, IT operations have greater control over the privacy of their data, which may include sensitive consumer or product information. The on-premises deployment of this solution also offers a potential economic advantage in its ability to avoid cloud storage costs when storing large capacities of video data. It additionally avoids the high networking requirements of uploading many video streams to the cloud.
Specifications of the Dell PowerEdge servers used in this PoC can be found in Figure 5
PowerEdge R7615 |
| |
Device Name |
| Dell PowerEdge R7615 |
CPU | Model Name | AMD EPYC 9354P 32-Core Processor |
Number Of Cores per Socket | 32 | |
Number Of Sockets | 1 | |
Memory | Size | 768 GB |
Storage | Size | 1 TB |
Network |
| Broadcom NetXtreme-E BCM57508 |
OS | Name | Ubuntu 22.04.3 LTS |
Kernel | 5.15.0-86-generic |
Figure 5: Dell PowerEdge Server Details
Key Highlights for IT Operations
|
Retail Solution Performance Observations
A key performance metric for the retail inventory management reference solution is the throughput of images per second as they are streamed by the in-store video cameras, decoded, and inferenced by the video pipeline. Video data is a common source for AI applications in the retail market, due to the prevalence of existing cameras deployed in stores, and the value of information that can be obtained by the video data. Because of this, the throughput performance insights gained from this PoC can translate to additional retail solutions that rely on image processing.
To examine the performance of the 32 core AMD EPYC 9354P processor for data capture and inferencing, the video pipeline was tested both with and without ZenDNN performance tuning, as well as with core pinning and node pinning. ZenDNN is a library that optimizes the performance of AMD processors for deep learning inferencing applications. The node pinning and core pinning are techniques offer optimization by binding processes to specific NUMA nodes or cores. The tests were run with up to 64 processes running on a 32 core server. The results of this testing can be seen in Figure 6.
Figure 6: Throughput Performance
The performance results demonstrate that the use of ZenDNN with node pinning can provide a dramatic increase in throughput, with mostly lower CPU utilization. On average, ZenDNN with node pinning achieved a throughput increase of approximately 1.5x. Further throughput increases were additionally achieved by utilizing core pinning. Full results can be seen in Figure 7.
Processes | Throughput Images/sec - ZenDNN | Throughput Images/sec - ZenDNN OFF |
| |||
Core Pinning | Node pinning | CPU utilization | Default | CPU utilization | Difference ZenD- NN(Node pinning) vs Default | |
1 | 29.86 | 31.72 | 7.808695652 | 25.06 | 10.75217391 | 1.27 |
8 | 195.7 | 188.26 | 46.27717391 | 125.02 | 59.36684783 | 1.51 |
16 | 305.06 | 264.24 | 62.7548913 | 176.99 | 75.2388587 | 1.49 |
32 | 389.1 | 347.58 | 78.978125 | 204.98 | 83.00978261 | 1.7 |
64 | 460.88 | 392.32 | 93.09952446 | 214.43 | 91.55903533 | 1.83 |
Figure 7: Video Pipeline Throughput Test
The performance gains achieved with ZenDNN, core pinning, and node pinning demonstrate the ability to optimize CPUs for AI applications. Commonly, computationally demanding AI processes, such as the computer vision and object detection utilized in this PoC, are expected to require GPUs. Hardware alone, however, is not the only component that affects performance. Software such as ZenDNN plays a key role in optimizing the performance of the chosen hardware, as does configuration details such as utilizing core pinning or node pinning. By utilizing these configurations, organizations can achieve AI applications that meet their performance needs with a CPU-based solution utilizing readily available hardware.
The PoC solution was additionally tested with an increasing number of video streams to assess the bandwidth of the networked video pipeline and visualization service. 1080p video was streamed to the video pipeline where it was decoded and inferenced. It was then transmitted and received by the visualization pipeline to be encoded and shared. The number of video streams was increased incrementally between 1 and 20 which resulted in an increasing bandwidth utilization. The bandwidth scaled from an average utilization of 1.65 Gbits/s and a max utilization of 3.4 Gbits/s with 1 stream, to an average utilization of 13.9 Gbits/s and a max utilization of 27.4 Gbits/s with 20 streams. An overview of the results can be seen in Figure 8.
Figure 8: Inventory Management System Bandwidth
Notably, the bandwidth does not increase linearly in relation to the number of streams, allowing the solution to scale as additional streams are needed. As the number of streams increases, however, the solution does experience a decrease in frames-per-second. While frames-per-second decreases, the overall utility of the solution is not significantly impacted. Higher frame rates are of greater importance when considering video with large amounts of motion, or when viewing quality is a major priority. In this particular solution, lower frame rates are acceptable as the focus is stationary store shelves, and real time viewing is not the primary use case. Full results of testing the networked solution, including both bandwidth utilization and frames per second, can be seen in Figure 9.
Number of Streams | AVG FPS / Stream | Throughput (FPS) | Avg Bandwidth Util (Gbits/s) | Max Bandwidth Util (Gbits/s) | Avg CPU Util (%) | Avg Memory Util (GB) |
1 | 31.14 | 31.14 | 1.65 | 3.4 | 12.61 | 6.5 |
2 | 30.92 | 61.84 | 3.2 | 6.7 | 21.8 | 7.27 |
4 | 28.78 | 115.12 | 6.2 | 12.2 | 41.38 | 9.2 |
8 | 22.17 | 177.36 | 9.86 | 20.5 | 65.06 | 13.9 |
10 | 20.53 | 205.3 | 11.2 | 22.4 | 73.18 | 16.4 |
12 | 18.8 | 225.6 | 12.1 | 24.7 | 78.76 | 18.2 |
16 | 13.97 | 223.52 | 12.6 | 25.6 | 81.39 | 22.2 |
20 | 11.7 | 234 | 13.9 | 27.4 | 84.1 | 26.7 |
Figure 9: Inventory Management System Bandwidth Test
The results of this performance testing demonstrate that the bandwidth of the networked servers is capable of scaling alongside more demanding video requirements. The separation of the video pipeline and the visualization service onto distinct servers allows the architecture to independently scale the compute resources for the two services. To capitalize on this architecture however, the networking between the servers must be capable of providing adequate bandwidth between the services. To do so, the PoC solution utilizes Broadcom BCM57508 NetXtreme-E Ethernet controllers capable of supporting up to 200GbE. By utilizing a modular architecture that’s connected with scalable, high bandwidth networking, the retail inventory management PoC provides a flexible starting point for retail organizations to scale to their individual needs, including the number of video streams, FPS requirements, and additional application logic.
Final Thoughts
With the rapid development of AI technology, the retail market presents many opportunities to deploy valuable new AI-powered applications. With the broad range of value that AI can bring to retail environments, both in improving CX and optimizing store operations, retail organizations should look to be proactive in adopting the emerging technology.
As a new technology, there are many unknowns and misconceptions for those in IT who may be unfamiliar with AI deployments, complicating and delaying new AI applications. A common challenge faced by IT is the expectation that AI applications will require specialized hardware solutions that are inaccessible. The AI-powered retail inventory management solution outlined in this paper serves as a demonstration of a broadly applicable AI solution for retail that can be deployed on off-the-shelf hardware solutions. The Dell hardware solutions used in the PoC deployment were demonstrated to handle the high-bandwidth video requirements as well as the AI modeling and inferencing requirements without the use of purpose-built accelerators, GPUs, or custom hardware.
The PoC solution outlined in this paper additionally serves as a reference for retail organizations to quickly deploy their own inventory management solution. While the solution discussed in this paper is limited to a PoC, it was designed with scalability in mind for organizations to further develop and scale a solution for their needs.
The use of an AI-powered inventory management system can provide real value and cost savings to organizations by avoiding over- or under-stocking products. By using readily available hardware and reference solutions, the barrier of entry for deploying such an AI solution is dramatically lowered, allowing retail organizations to achieve quicker deployments of new AI applications and quicker time to value.
CONTRIBUTORS
Mitch Lewis
Research Analyst | The Futurum Group
PUBLISHER Daniel Newman
CEO | The Futurum Group
INQUIRIES
Contact us if you would like to discuss this report and The Futurum Group will respond promptly.
CITATIONS
This paper can be cited by accredited press and analysts, but must be cited in-context, displaying author’s name, author’s title, and “The Futurum Group.” Non-press and non-analysts must receive prior written permission by The Futurum Group for any citations.
LICENSING
This document, including any supporting materials, is owned by The Futurum Group. This publication may not be reproduced, distributed, or shared in any form without the prior written permission of The Futurum Group.
DISCLOSURES
The Futurum Group provides research, analysis, advising, and consulting to many high-tech companies, including those mentioned in this paper. No employees at the firm hold any equity positions with any companies cited in this document.
ABOUT THE FUTURUM GROUP
The Futurum Group is an independent research, analysis, and advisory firm, focused on digital innovation and market-disrupting technologies and trends. Every day our analysts, researchers, and advisors help business leaders from around the world anticipate tectonic shifts in their industries and leverage disruptive innovation to either gain or maintain a competitive advantage in their markets.
![Post thumbnail](https://cdn-prod.scdn6.secure.raxcdn.com/static/media/527fe3b7-1deb-4ab7-b00a-b51b3e2649fa.jpeg?_cb=1709920955.8132777)
Dell POC for Scalable and Heterogeneous Gen-AI Platform
Fri, 08 Mar 2024 18:35:58 -0000
|Read Time: 0 minutes
Introduction
As part of Dell’s ongoing efforts to help make industry leading AI workflows available to their clients, this paper outlines a scalable AI concept that can utilize heterogeneous hardware components. The featured Proof of Concept (PoC) showcases a Generative AI Large Language Model (LLM) in active production, capable of functioning across diverse hardware systems.
Currently, most AI offerings are highly customized and designed to operate with specific hardware, either a particular vendor's CPUs or a specialized hardware accelerator such as a GPU. Although the operational stacks in use vary across different operational environments, they maintain a core similarity and adapt to each specific hardware requirement.
Today, the conversation around Generative-AI LLMs often revolves around their training and the methods for enhancing their capabilities. However, the true value of AI comes to light when we deploy it in production. This PoC focuses on the application of generative AI models to generate useful results. Here, the term 'inferencing' is used to describe the process of extracting results from an AI application.
As companies transition AI projects from research to production, data privacy and security emerge as crucial considerations. Utilizing corporate IT-managed equipment and AI stacks, firms ensure the necessary safeguards are in place to protect sensitive corporate data. They effectively manage and control their AI applications, including security and data privacy, by deploying AI applications on industry-standard Dell servers within privately managed facilities.
Multiple PoC examples on Dell PowerEdge hardware, offering support for both Intel and AMD CPUs, as well as Nvidia and AMD GPU accelerators. These configurations showcase a broad range of performance options for production inferencing deployments. Following our previous Dell AI Proof of Concept,[1] which examined the use of distributed fine-tuning to personalize an AI application, this PoC can serve as the subsequent step, transforming a trained model into one that is ready for production use.
Designed to be industry-agnostic, this PoC provides an example of how we can create a general-purpose generative AI solution that can utilize a variety of hardware options to meet specific Gen-AI application requirements.
In this Proof of Concept, we investigate the ability to perform scale-out inferencing for production and to utilize a similar inferencing software stack across heterogeneous CPU and GPU systems to accommodate different production requirements. The PoC highlights the following:
- A single CPU based system can support multiple, simultaneous, real-time sessions
- GPU augmented clusters can support hundreds of simultaneous, real-time sessions
- A common AI inferencing software architecture is used across heterogenous hardware
Futurum Group Comment: The novel aspect of this proof of concept is the ability to operate across different hardware types, including Intel and AMD CPUs along with support for both Nvidia and AMD GPUs. By utilizing a common inferencing framework, organizations are able to choose the most appropriate hardware deployment for each application’s requirements. This unique approach helps reduce the extensive customization required by AI practitioners, while also helping IT operations to standardize on common Dell servers, storage and networking components for their production AI deployments. |
Distributed Inferencing PoC Highlights
The inferencing examples include both single node CPU only systems, multi-node CPU clusters, along with single node and clusters of GPU augmented systems. Across this range of hardware options, the resulting generative AI application provides a broad range of performance, ranging from the ability to support several interactive query and response streams on a CPU, up to the highest performing example supporting thousands of queries utilizing a 3-node cluster with GPU cards to accelerate performance.
The objective of this PoC was to evaluate the scalability of production deployments of Generative-AI LLMs on various hardware configurations. Evaluations included deployment on CPU only, as well as GPU assisted configurations. Additionally, the ability to scale inferencing by distributing the workload across multiple nodes of a cluster were investigated. Various metrics were captured in order to characterize the performance and scaling, including the total throughput rate of various solutions, the latency or delay in obtaining results, along with the utilization rates of key hardware elements.
The examples included between one to three Dell PowerEdge servers with Broadcom NICs, with additional GPU acceleration provided in some cases by either AMD or Nvidia GPUs. Each cluster configuration was connected using Broadcom NICs to a Dell Ethernet switch for distributed inferencing. Each PoC uses one or more PowerEdge servers, with some examples also using GPUs. Dell PowerEdge Servers used included a Dell XE8545, a Dell XE9680 and a Dell R760XA. Each Dell PowerEdge system also included a Broadcom network interface (NIC) for all internode communications connected via a Dell PowerSwitch.
Shown in Figure 1 below is a 3-node example that includes the hardware and general software stack.
Figure 1: General, Scale-Out AI Inferencing Stack (Source: Scalers.AI)
There are several important aspects of the architecture utilized, enabling organizations to customize and deploy generative AI applications in their choice of colocation or on-premises data center. These include:
- Dell PowerEdge Sixteenth Gen Servers, with 4th generation CPUs and PCIe Gen 5 connectivity
- Broadcom NetXtreme BCM57508 NICs with up to 200 Gb/s per ethernet port
- Dell PowerSwitch Ethernet switches Z line support up to 400 Gb/s connectivity
This PoC demonstrating both heterogeneous and distributed inferencing of LLMs provides multiple advantages compared to typical inferencing solutions:
- Enhanced Scalability: Distributed inferencing enables the use of multiple nodes to scale the solution to the desired performance levels.
- Increased Throughput: By distributing the inferencing processes across multiple nodes, the overall throughput increases.
- Increased Performance: The speed of generated results may be an important consideration, by supporting both CPU and GPU inferencing, the appropriate hardware can be selected.
- Increased Efficiency: Providing a choice of using CPU or GPUs and the number of nodes enables organizations to align the solutions capabilities with their application requirements.
- Increased Reliability: With distributed inferencing, even if one node fails, the others can continue to function, ensuring that the LLM remains operational. This redundancy enhances the reliability of the system
Although each of the capabilities outlined above are related, certain considerations may be more important than others for specific deployments. Perhaps more importantly, this PoC demonstrates the ability to stand up multiple production deployments, using a consistent set of software that can support multiple deployment scenarios, ranging from a single user up to thousands of simultaneous requests. In this PoC, there are multiple hardware solution deployment examples, summarized as follows:
- CPU based inferencing using both AMD and Intel CPUs, scaled from 1 to 2 nodes
- GPU based inferencing using Nvidia and AMD GPUs, scaled from 1 to 3 nodes
For the GPU based configurations, a three node 12-GPU configuration achieved nearly 3,000 words per second in total output generation. For the scale-out configurations, inter-node communications were an important aspect of the solution. Each configuration utilized a Broadcom BCM-57508 Ethernet card enabling high-speed and low latency communications. Broadcom’s 57508 NICs allow data to be loaded directly into accelerators from storage and peers, without incurring extra CPU or memory overhead.
Futurum Group Comment: By using a scale out inferencing solution leveraging industry standard Dell servers, networking and optional GPU accelerators provides a highly adaptable reference that can be deployed as an edge solution where few inferencing sessions are required, up to enterprise deployments supporting hundreds of simultaneous inferencing outputs. |
Evaluating Solution Performance
In order to compare the performance of the different examples, it is important to understand some of the most important aspects of commonly used to measure LLM inferencing. These include the concept of a token, which typically consists of a group of characters, which are groupings of letters, with larger words comprised of multiple tokens. Currently, there is no standard token size utilized across LLM models, although each LLM typically utilizes common token sizes. Each of the PoCs utilize the same LLM and tokenizer, resulting in a common ratio of tokens to words across the examples. Another common metric is that of a request, which is essentially the input provided to the LLM and may also be called a query.
A common method of improving the overall efficiency of the system is to batch requests, or submit multiple requests simultaneously, which improves the total throughput. While batching requests increases total throughput, it comes at the cost of increasing the latency of individual requests. In practice, batch sizes and individual query response delays must be balanced to provide the response throughput and latencies that best meet a particular application’s needs.
Other factors to consider include the size of the base model utilized, typically expressed in billions of parameters, such as Mistral-7B (denoting 7 billion parameters), or in this instance, Llama2-70B, indicating that the base model utilized 70 billion parameters. Model parameter sizes are directly correlated to the necessary hardware requirements to run them.
Performance testing was performed to capture important aspects of each configuration, with the following metrics collected:
- Requests per Second (RPS): A measure of total throughput, or total requests processed per second
- Token Throughput: Designed to gauge the LLMs performance using token processing rate
- Request Latency: Reports the amount of delay (latency) for the complete response, measured in seconds, and for individual tokens, measured in milli-seconds.
- Hardware Metrics: These include CPU, GPU, Network and Memory utilization rates, which can help determine when resources are becoming overloaded, and further splitting or “sharding” of a model across additional resources is necessary.
Note: The full testing details are provided in the Appendix.
Testing evaluated the following aspects and use cases:
- Effects of scaling for interactive use cases, and batch use cases
- Scaling from 1 to 3 nodes, for GPU configurations, using 4, 8, 12 and 16 total GPUs
- Scaling from 1 to 3 nodes for CPU only configurations (using 112, 224 and 448 total CPU cores)
- For GPU configuration, the effect of moderate batch sizes (32) vs. large batching (256)
- Note: for CPU configurations, the batch size was always 1, meaning a single request per instance
We have broadly stated that two different use cases were tested, interactive and batch. An interactive use case may be considered an interactive chat agent, where a user is interacting with the inferencing results and expects to experience good performance. We subsequently define what constitutes “good performance” for an interactive user. An additional use case could be batch processing of large numbers of documents, or other scenarios where a user is not directly interacting with the inferencing application, and hence there is no requirement for “good interactive performance”.
As noted, for GPU configurations, two different batch sizes per instance were used, either 32 or 256. Interactive use of an LLM application may be uses such as chatbots, where small delays (i.e. low latency) is the primary consideration, and total throughput is a secondary consideration. Another use case is that of processing documents for analysis or summarization. In this instance, total throughput is the most important objective and the latency of any one process is inconsequential. For this case, the batch operation would be more appropriate, in order to maximize hardware utilization and total processing throughput.
Interactive Performance
For interactive performance, the rate of text generation should ideally match, or exceed the users reading or comprehension rate. Also, each additional word output should be created with relatively small delays. According to The Futurum Group’s analysis of reading rates, 200 words per minute can be considered a relatively fast rate for comprehending unseen, non-fiction text. Using this as a guideline results in a rate of 3.33 words per second.
- 200 wpm / 60 sec / minute = 3.33 words per second
Moreover, we will utilize a rate of 3.33 wps as the desired minimum generation rate for assessing the ability to meet the needs of a single interactive user. In terms of latency, 1 over 3.33, or 300 milliseconds would be considered an appropriate maximum delay threshold.
Note: For Figures 2 – 5, each utilizes two axes, the primary (left) vertical axis represents the throughput for the bars in words per second. The second (right) vertical axis represents the 95th percentile of latency results for each word generated.
In Figure 2 below, we show the total throughput of 3 different CPU configurations, along with the associated per word latency. As seen, a CPU only example can support over 40 words per second, significantly greater than the 3.33 word per second rate required for good interactive performance, while maintaining a latency of 152 ms., well under 300 ms.
Figure 2: Interactive Inferencing Performance for CPUs (Source: Futurum Group)
Using a rate of 3.33 words / sec., we can see that two system, each with 224 CPU cores can support inferencing of up to 12 simultaneous sessions.
- Calculated as: 40 wps / 3.33 wps / session = 12 simultaneous sessions.
Futurum Group Comment: It is often expected that all generative AI applications require the use of GPUs in order to support real-time deployments. As evidenced by the testing performed, it can be seen that a single system can support multiple, simultaneous sessions, and by adding a second system, performance scales linearly, doubling from 20 words per second up to more than 40 words per second. Moreover, for smaller deployments, a single CPU based system supporting inferencing may be sufficient. |
In Figure 3 below, we show the total throughput of 4 different configurations, along with the associated per word latency. As seen, even at the rate of 1,246 words per second, latency remains at 100 ms., well below our 300 ms. threshold.
Figure 3: Interactive Inferencing Performance for GPUs (Source: Futurum Group)
Again, using 3.33 words / sec., each example can support a large number of interactive sessions:
- 1 node + 4 GPUs: 414 wps / 3.33 wps / session = 124 simultaneous sessions
- 2 nodes + 8 GPUs: 782 wps / 3.33 wps / session = 235 simultaneous sessions
- 2 nodes + 12 GPUs: 1,035 wps / 3.33 wps / session = 311 simultaneous sessions
- 3 nodes + 16 GPUs: 1,246 wps / 3.33 wps / session = 374 simultaneous sessions
Futurum Group Comment: Clearly, the GPU based results significantly exceed those of the CPU based deployment examples. In these examples, we can see that once again, performance scales well, although not quite linearly. Perhaps more importantly, as additional nodes are added, the latency does not increase above 100 ms., which is well below our established desired threshold. Additionally, the inferencing software stack was very similar to the CPU only stack, with the addition of Nvidia libraries in place of Intel CPU libraries. |
Batch Processing Performance
Inferencing of LLMs becomes memory bound as the model size increases. For larger models such as Llama2-70B, memory bandwidth, between either the CPU and main memory, or GPU and GPU memory is the primary bottleneck. By batching requests, multiple processes may be processed by the GPU or CPU without loading new data into memory, thereby improving the overall efficiency significantly.
Having an inference serving system that can operate at large batch sizes is critical for cost efficiency, and for large models like Llama2-70B the best cost/performance occurs at large batch sizes.
In Figure 4 below we show the throughput capabilities of the same hardware configuration used in Figure 3, but this time with a larger batch size of 256.
Figure 4: Batch Inferencing Performance for GPUs (Source: The Futurum Group)
For this example, we would not claim the ability to support interactive sessions. Rather the primary consideration is the total throughput rate, shown in words per second. By increasing the batch size by a factor of 4X (from 32 to 256), the total throughput more than doubles, along with a significant increase in the per word latency, making this deployment appropriate for offline, or non-interactive scenarios.
Futurum Group Comment: Utilizing the exact same inferencing software stack, and hardware deployment, we can show that for batch processing of AI, the PoC example is able to achieve rates up to nearly 3,000 words per second. |
Comparison of Batch vs. Interactive
As described previously, we utilized a total throughput rate of 200 words per minute, or 3.33 words per second, which yields a maximum delay of 300 ms per word as a level that would produce acceptable interactive performance. In Figure 4 below, we compare the throughput and associated latency of the “interactive” configuration to the “batch” configuration.
Figure 5: Comparison of Interactive vs. Batch Inferencing on GPUs (Source: The Futurum Group)
As seen above, while the total throughput, measured in words per second increases by 2.4X, the latency of individual word output slows substantially, by a factor of 6X. It should be noted that in both cases batching was utilized. The batch size of the “interactive” results was set to 32, while the batch size of the “batch” results utilized a setting of 256. The “interactive” label was applied to the lower results, due to the fact that the latency delay of 100 ms. was significantly below the threshold of 300 ms. for typical interactive use. In summary:
- Throughput increase of 2.4X (1,246 to 2,962) for total throughput, measured in words per second
- Per word delays increased 6X (100 ms. to 604 ms.) measured as latency in milli-seconds
These results highlight that total throughput can be improved, albeit at the expense of interactive performance, with individual words requiring over 600 ms (sixth tenths of a second) when the larger batch size of 256 was used. With this setting, the latency significantly exceeded the threshold of what is considered acceptable for interactive use, where a latency of 300 ms would be acceptable.
Highlights for IT Operations
While terms such as tokens per second, and token latency have relevancy to AI practitioners, these are not particularly useful terms for IT professionals or users attempting to interact with generative LLM models. Moreover, we have translated these terms into more meaningful terminology that can help IT operations correctly size the hardware requirements to match expected usage. In particular, for interactive sessions requiring a rate of 200 words per second, and maximum delay of 300 ms. per word, we can then translate a total word per second throughput, into simultaneous streams. By using a rate of 3.33 words per second as the minimum per interactive session, we can determine the number of interactive sessions supported at a certain throughput and latency levels.
Nodes | Total Cores | Words /sec. | Word. Lat. (ms) | # Sessions |
1 | 112 | 11.45 | 111 | 3.4 |
1 | 224 | 20.17 | 140 | 6.1 |
2 | 448 | 40.75 | 152 | 12.2 |
Table 1: Interactive Inferencing Sessions using CPUs (Source: The Futurum Group)
Nodes | Total GPUs | Words /sec. | Word. Lat. (ms) | # Sessions |
1 | 4 | 414 | 76 | 124 |
2 | 8 | 782 | 100 | 235 |
2 | 12 | 1035 | 100 | 310 |
3 | 16 | 1246 | 100 | 374 |
Table 2: Interactive Inferencing Sessions using GPUs (Source: The Futurum Group)
The ability to scale inferencing solutions is important, as outlined previously. Additionally, perhaps the most unique aspect of this PoC is the ability to support operating an AI inferencing stack across both CPU only and GPU enhanced hardware architectures, using an optimized inferencing stacks for each hardware type.
For environments requiring only a few simultaneous inferencing sessions, it is possible to meet these needs with CPU only deployments, even when using a larger LLM model such as Llama2-70B utilized during testing. A current Dell PowerEdge server with a 4th generation processor can support up to 3 simultaneous interactive inferencing sessions per server, with a 3-node CPU cluster able to support up to 12 simultaneous sessions, at a total rate of 40 words per second across all three systems.
Use cases that require higher throughput, or the ability to support a greater number of simultaneous inferencing sessions can utilize a single GPU based PowerEdge server with 4 GPUs, which was found to support up to 124 simultaneous interactive sessions. Scaling beyond this, a 3-node, 16 GPU system was able to support 373 simultaneous inferencing sessions at a rate of 200 words per minute, for a total throughput of 1,245 words per second.
Highlights for AI Practitioners
A key aspect of the PoC is the software stack that helps provide a platform for AI deployments, enabling scale-out infrastructure to significantly increase content creation rates. Importantly, this AI Platform as a Service architecture was built using Dell and Broadcom hardware components, coupled with cloud native components to enable containerized software platform with open licensing to reduce deployment friction and reduce cost.
At a high-level, the inferencing stack is comprised of Ray-Serve, under-pined by vLLM to improve memory utilization during inferencing, including Hugging Face Optimum libraries and ONNX. Additionally, specific libraries were used to further enhance performance, including ROCm for AMD CPUs and GPUs, BigDL for Intel CPUs and CUDA for Nvidia GPUs. Other AI frameworks include the use of PyTorch within each container, along with Kubernetes and KubeRay for distributed cluster management. Shown in Figure 5 is a high-level architecture of the inferencing stack used for all hardware deployments.
Figure 6: Distributed Inferencing Framework for Heterogeneous Hardware (Source: Scalers.AI)
Variations of the depicted inferencing framework were utilized for the four different hardware types, (AMD CPU, Intel CPU, AMD GPU and Nvidia GPU). The specifics of each of these are provided in the Appendix, in Figures A1-A4.
Note: It is important to recognize that while both Intel and AMD CPUs were verified and tested, only the Intel CPU results are presented here. This is for several reasons, including the fact that comparing CPU performance was not an objective of this PoC. If competing systems performance was provided, the focus would artificially become more about comparing CPU results than showcasing the ability to run across different CPU types. Similarly, while both AMD and Nvidia GPU inferencing was verified, only the Nvidia GPU results are presented, to maintain the focus on the PoC capabilities rather than comparing different GPU vendors performance.
Final Thoughts
As artificial intelligence, and in particular Generative-AI matures, companies are seeking ways to leverage this new technology to provide advantages for their firms, helping to improve efficiency and other measures of user satisfaction. GenAI based Large Language Models are quickly showing their ability to augment some applications focused both on empowering internal users with additional knowledge and insights, and is also becoming increasingly useful for assisting clients, via chatbots or other similar interfaces. However, organizations often have concerns about becoming tied to proprietary, or cloud-based solutions, due to their privacy concerns, lack of transparency or potential vendor lock in.
As part of Dell’s continuing efforts to democratize AI solutions, this proof-of-concept outlines specifically how organizations can build, deploy and operate production use of generative AI models using industry standard Dell servers. In particular, the scale-out PoC detailed in this paper showcases the ability to scale a solution efficiently from supporting a few simultaneous interactive users, up to a deployment supporting hundreds of simultaneous inferencing sessions simultaneously using as few as 3 Dell PowerEdge servers augmented by Nvidia GPUs. In an offline, or batch processing use, the same hardware example can support a throughput of nearly 3,000 words per second when processing multiple documents.
Critically, all the examples leverage a common AI framework, consisting of minimal K3s Kubernetes deployments, along with the Ray framework for distributed processing and vLLM to improve distributed inferencing performance. The outlined PoC utilizes the Hugging Face repository and libraries, along with hardware specific optimizations for each specific deployment of CPU or GPU type. By using a common framework, AI practitioners are better able to focus on enhancing the accuracy of the models and improved training methodologies, rather than trying to debug multiple solution stacks. Likewise, IT operations staff can utilize standard hardware, along with common IT technologies such as Kubernetes running on standard Linux distributions.
References
[1] Futurum Group Labs: Dell and Broadcom Release Scale-Out Training for Large Language Models
Appendix
Due to the fact that LLM inferencing is often memory bound, and due to the manner in which LLMs iteratively generate output, it is possible to optimize performance by batching input. By batching input, more queries are present within the GPU memory card, for the LLM to generate output leveraging the GPU processing. In this way, the primary bottleneck of moving data into and out of GPU memory is reduced per output generated, thereby increasing the throughput, with some increase in latency per request. These optimizations include the use of vLLM, Ray-Serve and Hugging Face Optimizations available through their Optimum inferencing models.
Moreover, continuous batching was utilized, in order to increase throughput with the vLLM library, which helps to manage memory efficiently. Two batch sizes were used with GPU configurations, 32 to provide a lower latency with good throughput, and a larger batch size of 256 to provide the highest throughput for non-interactive use cases where latency was not a concern.
Note: other metrics such as time to first token (TTFT) may be gathered; however, in our testing, the TTFT was not deemed to be a critical element for analysis.
3. Test Scenarios
The Llama 2 70B Chat HF model is loaded with a tensor parallelism of 4 GPUs. A 70B model (Float 32 precision) requires ~260 GB of GPU memory to load the model. Based on the model weight GPU memory requirement and inference requirements, we recommend using 4x80GB GPUs to load a single Llama 2 70B Chat model.
The Llama 2 70B Chat model with bfloat16 precision was used for all test configurations.
AI Inferencing Stack Details
In the figures below, we highlight the various specific stacks utilized for each hardware deployment.
Figure A1: Nvidia GPU inferencing stack (Source: Scalers.AI)
Figure A2: AMD GPU inferencing stack (Source: Scalers.AI)
Figure A3: Intel CPU inferencing stack (Source: Scalers.AI)
Figure A4: AMD CPU inferencing stack (Source: Scalers.AI)
Single Node Inferencing
The below table describes the single node inferencing Kubernetes deployment configuration with 4 GPUs (1 replica).
Device | Node Type | GPU | GPU Count | CPU Cores | Memory | Disk |
Dell PowerEdge XE9680 | Head | - | - | 160 | 300 GB | 1 TB |
Dell PowerEdge XE8545 | Worker | NVIDIA A100 SXM 80GB | 4 | 160 | 300 GB | 1 TB |
Table A1: Interactive Inferencing Sessions using GPUs (Source: Futurum Group)
Two Node Inferencing
Scenario 1: 8 GPUs, 2 Replicas, shown below.
Device | Node Type | GPU | GPU Count | CPU Cores | Memory | Disk |
Dell PowerEdge XE9680 | Head+ Worker | NVIDIA A100 SXM 80GB | 4 | 160 | 300 GB | 1 TB |
Dell PowerEdge XE8545 | Worker | NVIDIA A100 SXM 80GB | 4 | 160 | 300 GB | 1 TB |
Table A2: Interactive Inferencing Sessions using GPUs (Source: Futurum Group)
Scenario 2: 12 GPUs, 3 Replicas
Device | Node Type | GPU | GPU Count | CPU Cores | Memory | Disk |
Dell PowerEdge XE9680 | Head+ Worker | NVIDIA A100 SXM 80GB | 8 | 160 | 300 GB | 1 TB |
Dell PowerEdge XE8545 | Worker | NVIDIA A100 SXM 80GB | 4 | 160 | 300 GB | 1 TB |
Table A3: Interactive Inferencing Sessions using GPUs (Source: Futurum Group)
Three Node Inferencing
The below table describes the two node inferencing hardware configuration with 16 GPUs(4 replicas).
Device | Node Type | GPU | GPU Count | CPU Cores | Memory | Disk |
Dell PowerEdge XE9680 | Head+ Worker | NVIDIA A100 SXM 80GB | 8 | 160 | 300 GB | 1 TB |
Dell PowerEdge XE8545 | Worker | NVIDIA A100 SXM 80GB | 4 | 160 | 300 GB | 1 TB |
Dell PowerEdge R760xa | Worker | NVIDIA H100 PCIe 80GB | 4 | 160 | 300 GB | 1 TB |
Table A4: Interactive Inferencing Sessions using GPUs (Source: Futurum Group)
Test Workload Configuration
The workload consists of a set of 1000+ prompts passed randomly for each test [ML1] with different concurrent requests. The concurrent requests are generated by Locust tool.
The inference configuration is as below
- Input token length: 14 to 40.
- Output token length: 256
- Temperature: 1
The tests were run with two different batch sizes per replica - 32 and 256.
Test Metrics
The below are the metrics measured for each tests
Metric | Explanation |
Requests Per Second (RPS) | Evaluate system throughput, measuring requests processed per second. |
Total Token Throughput (tokens/s) | Quantify language model efficiency by assessing token processing rate. |
Request Latency (P50, P95, P99) | Gauge system responsiveness through different latency percentiles. |
Average CPU, Memory, GPU Utilization | Assess system resource usage, including CPU, memory, and GPU. |
Network Bandwidth (Average, Maximum) | Measure efficiency in data transfer with average and maximum network bandwidth. |
Table A5: Interactive Inferencing Sessions using GPUs (Source: Futurum Group)
Performance Results
Scalability results for batch size of 32 per replica
Inference Nodes | Devices | GPUs | Concurrent Requests | RPS | Tokens/s | P95 Latency(s) | P95 Token Latency(ms) | |
Single Node | Dell PowerEdge XE8545 | 4xNVIDIA A100 SXM 80GB | 32 | 2.7 | 621.4 | 13 | 50.78 | |
Two Nodes | Dell PowerEdge XE9680(4 GPUs) | 8xNVIDIA A100 SXM 80GB | 64 | 4.8 | 1172.63 | 17 | 66.41 | |
Two Nodes | Dell PowerEdge XE9680 | 12xNVIDIA A100 SXM 80GB | 96 | 6.8 | 1551.94 | 17 | 66.4 | |
Three Nodes | Dell PowerEdge XE9680 Dell PowerEdge R760xa | 12xNVIDIA A100 SXM 80GB 4xNVIDIA H100 PCIe 80GB | 128 | 8.3 | 1868.76 | 17 | 66.4 |
Table A6: Interactive Inferencing Sessions using GPUs (Source: Futurum Group)
Scalability results for batch size of 256 per replica
Inference Nodes | Devices | GPUs | Concurrent Requests | RPS | Throughput | P95 Latency | P95 Token Latency(ms) | |
Single Node | Dell PowerEdge XE8545 | 4xNVIDIA A100 SXM 80GB | 256 | 6.4 | 1475.64 | 45 | 175.78 | |
Two Nodes | Dell PowerEdge XE9680(4 GPUs) | 8xNVIDIA A100 SXM 80GB | 512 | 10.3 | 2542.32 | 61 | 238.28 | |
Two Nodes | Dell PowerEdge XE9680 | 12xNVIDIA A100 SXM 80GB | 768 | 14.5 | 3222.89 | 64 | 250 | |
Three Nodes | Dell PowerEdge XE9680 Dell PowerEdge R760xa | 12xNVIDIA A100 SXM 80GB 4xNVIDIA H100 PCIe 80GB | 1024 | 17.5
| 4443.5 | 103 | 402.34 |
Table A7: Interactive Inferencing Sessions using GPUs (Source: Futurum Group)
For more detailed report refer: Dell Distributed Inference Data
Test Methodology for Distributed Inferencing on CPUs
Server | CPU | RAM | Disk |
Dell PowerEdge XE9680 | Intel(R) Xeon(R) Platinum 8480+ | 2 TB | 3 TB |
Dell PowerEdge R760xa | Intel(R) Xeon(R) Platinum 8480+ | 1 TB | 1 TB |
Table A8: Interactive Inferencing Sessions using CPUs (Source: Futurum Group)
Each server is networked to a Dell PowerSwitch Z9664F-ON through Broadcom BCM57508 NICs with 100 Gb/s bandwidth.
3. Test Scenarios
The Llama 2 7B Chat HF model is tested on CPU with int8 precision.
3.1 Single Node Inferencing
Three are two scenarios for single node inferencing
Scenario 1: 112 Cores, 1 Replicas
The below table describes the single node inferencing Kubernetes deployment configuration with 112 CPU Cores of Dell PowerEdge R760xa server (1 replica).
Device | Node Type | CPU Cores | Memory | Disk |
Dell PowerEdge XE9680 | Head | - | 500 GB | 1 TB |
Dell PowerEdge R760xa | Worker | 112 | 500 GB | 1 TB |
Table A9: Interactive Inferencing Sessions using GPUs (Source: Futurum Group)
Scenario 2: 224 Cores, 2 Replicas
The below table describes the single node inferencing Kubernetes deployment configuration with 224 CPU Cores of Dell PowerEdge R760xa server (2 replicas).
Device | Node Type | CPU Cores | Memory | Disk |
Dell PowerEdge XE9680 | Head | 10 | 500 GB | 1 TB |
Dell PowerEdge R760xa | Worker | 224 | 500 GB | 1 TB |
Table A10: Interactive Inferencing Sessions using GPUs (Source: Futurum Group)
3.2 Two Node Scenario
The below table describes the two-node inferencing hardware configuration with both the servers, 448 Cores of CPU (4 replicas).
Device | Node Type | CPU Cores | Memory | Disk |
Dell PowerEdge XE9680 | Head | 224 | 500 GB | 1 TB |
Dell PowerEdge R760xa | Worker | 224 | 500 GB | 1 TB |
Table A11: Interactive Inferencing Sessions using GPUs (Source: Futurum Group)
3. Test Workload Configuration
The workload consists of a set of 1000+ prompts passed randomly for each test with different concurrent requests. The concurrent requests are generated by Locust tool.
The inference configuration is as below
- Input token length: 14 to 40.
- Output token length: 256
- Temperature: 1
The tests were run with 1 batch size per replica.
5. Test Metrics
The below are the metrics measured for each tests
Metric | Explanation |
Requests Per Second (RPS) | Evaluate system throughput, measuring requests processed per second. |
Total Token Throughput (tokens/s) | Quantify language model efficiency by assessing token processing rate. |
Request Latency (P50, P95, P99) | Gauge system responsiveness through different latency percentiles. |
Average CPU, Memory | Assess system resource usage, including CPU, memory. |
Network Bandwidth (Average, Maximum) | Measure efficiency in data transfer with average and maximum network bandwidth. |
Table A12: Interactive Inferencing Sessions using GPUs (Source: Futurum Group)
4. Performance Reports
Inference Nodes | Devices | CPU Cores | Concurrent Requests | RPS | Throughput | P95 Latency(s) | P95 Token Latency(ms) | |
Single Node | Dell PowerEdge R760xa | Intel Xeon Platinum 8480+ (112 Cores) | 1 | 0.1 | 17.18 | 18 | 70.31 | |
Single Node | Dell PowerEdge R760xa | Intel Xeon Platinum 8480+ (224 Cores) | 2 | 0.1 | 30.26 | 21 | 82.03
| |
Two Nodes | Dell PowerEdge XE9680 | Intel Xeon Platinum 8480+ (448 Cores) | 4 | 0.3 | 61.13
| 23 | 89.84 |
Table A13: Interactive Inferencing Sessions using GPUs (Source: Futurum Group)
About The Futurum Group
The Futurum Group is dedicated to helping IT professionals and vendors create and implement strategies that make the most value of their storage and digital information. The Futurum Group services deliver in-depth, unbiased analysis on storage architectures, infrastructures, and management for IT professionals. Since 1997 The Futurum Group has provided services for thousands of end-users and vendor professionals through product and market evaluations, competitive analysis, and education.
Copyright 2024 The Futurum Group. All rights reserved.
No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying and recording, or stored in a database or retrieval system for any purpose without the express written consent of The Futurum Group. The information contained in this document is subject to change without notice. The Futurum Group assumes no responsibility for errors or omissions and makes no expressed or implied warranties in this document relating to the use or operation of the products described herein. In no event shall The Futurum Group be liable for any indirect, special, inconsequential, or incidental damages arising out of or associated with any aspect of this publication, even if advised of the possibility of such damages. All trademarks are the property of their respective companies.
This document was developed with funding from Dell Inc. and Broadcom. Although the document may utilize publicly available material from various vendors, including Dell, Broadcom and others, it does not necessarily reflect such vendors' positions on the issues addressed in this document.