Dell Technologies has been an active participant in the MLCommons™ Inference benchmark submission since day one. We have completed five rounds of inference submission.

This blog provides an overview of the latest results of MLPerf Inference v2.0 closed data center, closed data center power, closed edge, and closed edge power categories on Dell servers from our HPC & AI Innovation Lab. It shows optimal inference and power (performance per watt) performance for Dell GPU-based servers (DSS 8440, PowerEdge R750xa, PowerEdge XE2420, PowerEdge XE8545, and PowerEdge XR12). The previous blog about MLPerf Inference v1.1 performance results can be found here.

What is new?

• There were 3,800 performance results for this round compared to 1,800 performance results for v1.1. Additionally, 885 systems in v2.0 compared to 424 systems in v1.1 shows that there were more than twice the systems submitted for this round. 
• For the 3D U-Net benchmark, the dataset now used is the KiTs 2019 Kidney Tumor Segmentation set.
• Early stopping was introduced in this round to replace a deterministic minimum query count with a function that dynamically determines when further runs are not required to identify additional performance gain.

Results at a glance

 Dell Technologies submitted 167 results to the various categories. The Dell team made 86 submissions to the closed data center category, 28 submissions to the closed data center power category, and 53 submissions to the closed edge category. For the closed data center category, the Dell team submitted the second most results. In fact, Dell Technologies submitted results from 17 different system configurations with the NVIDIA TensorRT and NVIDIA Triton inference engines. Among these 17 configurations, the PowerEdge XE2420 server with T4 and A30 GPUs and the PowerEdge XR12 server with the A2 GPU were two new systems that have not been submitted before. Additionally, Dell Technologies submitted to the reintroduced Multiterm scenario. Only Dell Technologies submitted results for different host operating systems.

Noteworthy results

Noteworthy results include:

• The PowerEdge XE8545 and R750xa servers yield Number One results for performance per accelerator with NVIDIA A100 GPUs. The use cases for this top classification include Image Classification, Object Detection, Speech-to-text, Medical Imaging, Natural Language Processing, and Recommendation.
• The DSS 8440 server yields Number Two results for system performance for multiple benchmarks including Speech-to-text, Object Detection, Natural Language Processing, and Medical Image Segmentati on uses cases among all submissions.
• The PowerEdge R750xa server yields Number One results for the highest system performance for multiple benchmarks including Image Classification, Object Detection, Speech-to-text, Natural Language Processing, and Recommendation use cases among all the PCIe-based GPU servers.
• The PowerEdge XE8545 server yields Number One results for the lowest multistrand latency with NVIDIA Multi-Instance GPU (MIG) in the edge category for the Image Classification and Object Detection use cases.
• The PowerEdge XE2420 server yields Number One results for the highest T4 GPU inference results for the Image Classification, Speech-to-text, and Recommendation use cases.
• The PowerEdge XR12 server yields Number One results for the highest performance per watt with NVIDIA A2 GPU results in power for the Image Classification, Object Detection, Speech-to-text, Natural Language Processing, and Recommendation use cases.

MLPerf Inference v2.0 benchmark results

The following graphs highlight the performance metrics for the Server and Offline scenarios across the various benchmarks from MLCommons. Dell Technologies presents results as an method for our customers to identify options to suit their deep learning application demands. Additionally, this performance data serves as a reference point to enable sizing of deep learning clusters. Dell Technologies strives to submit as many results as possible to offer answers to ensure that customer questions are resolved.

For the Server scenario, the performance metric is queries per second (QPS). For the Offline scenario, the performance metric is Offline samples per second. In general, the metrics represent throughput, and a higher throughput indicates a better result. In the following graphs, the Y axis is an exponentially scaled axis representing throughput and the X axis represents the systems under test (SUTs) and their corresponding models. The SUTs are described in the appendix.

Figure 1 through Figure 6 show the per card performance of the various SUTs on the ResNet 50, BERT, SSD, 3dUnet, RNNT, and DLRM modes respectively in the Server and Offline scenarios:

Figure 1: MLPerf Inference v2.0 ResNet 50 per card results

Figure 2: MLPerf Inference v2.0 BERT default and high accuracy per card results

Figure 3: MLPerf Inference v2.0 SSD-ResNet 34 per card results

Figure 4: MLPerf Inference v2.0 3D U-Net per card results

Figure 5: MLPerf Inference v2.0 RNNT per card results

Figure 6: MLPerf Inference v2.0 DLRM default and high accuracy per card results

Observations

The results in this blog have been officially submitted to and accepted by the MLCommons organization. These results have passed compliance tests, been peer reviewed, and adhered to the constraints enforced by MLCommons. Customers and partners can reproduce our results by following steps to run MLPerf Inference v2.0 in its GitHub repository.

Submissions from Dell Technologies included approximately 140 performance results and 28 performance and power results. Across the various workload tasks including Image Classification, Object Detection, Medical Image Segmentation, Speech-to-text, Language Processing, and Recommendation, server performance from Dell Technologies was promising.

Dell servers performed with optimal performance and power results. They were configured with different GPUs such as:

• NVIDIA A30 Tensor Core GPU
• NVIDIA A100 (PCIe and SXM)
• NVIDIA T4 Tensor Core GPU
• NVIDIA A2 Tensor Core GPU, which is newly released

More information about performance for specific configurations that are not discussed in this blog can be found in the v1.1 or v1.0 results.

The submission included results from different inference backends such as NVIDIA TensorRT and NVIDIA Triton. The appendix provides a summary of the full hardware and software stacks.

Conclusion

This blog quantifies the performance of Dell servers in the MLPerf Inference v2.0 round of submission. Readers can use these results to make informed planning and purchasing decisions for their AI workload needs.

Appendix

Software stack

The NVIDIA Triton Inference Server is an open-source inferencing software tool that aids in the deployment and execution of AI models at scale in production. Triton not only works with all major frameworks but also with customizable backends, further enabling developers to focus on their AI development. It is a versatile tool because it supports any inference type and can be deployed on any platform including CPU, GPU, data center, cloud, or edge. Additionally, Triton supports the rapid and reliable deployment of AI models at scale by integrating well with Kubernetes, Kubeflow, Prometheus, and Grafana. Triton supports the HTTP/REST and GRPC protocols that allow remote clients to request inferencing for any model that the server manages.

The NVIDIA TensorRT SDK delivers high-performance deep learning inference that includes an inference optimizer and runtime. It is powered by CUDA and offers a unified solution to deploy on various platforms including edge or data center. TensorRT supports the major frameworks including PyTorch, TensorFlow, ONNX, and MATLAB. It can import models trained in these frameworks by using integrated parsers. For inference, TensorRT performs orders of magnitude faster than its CPU-only counterparts.

NVIDIA MIG can partition GPUs into several instances that extend compute resources among users. MIG enables predictable performance and maximum GPU use by running jobs simultaneously on the different instances with dedicated resources for compute, memory, and memory bandwidth.

SUT configuration

The following table describes the SUT from this round of data center inference submission:

Table 1: MLPerf Inference v2.0 system configurations for DSS 8440 and PowerEdge R750xa servers

Table 2: MLPerf Inference v2.0 system configurations for PowerEdge XE2420 servers

Table 3: MLPerf Inference v2.0 system configurations for PowerEdge XE8545 servers

Table 4: MLPerf Inference v2.0 system configurations for PowerEdge XR12 servers

Overview

The Dell PowerEdge R7525 server powered with 2nd Gen AMD EPYC processors was released as part of the Dell server portfolio. It is a 2U form factor rack-mountable server that is designed for HPC workloads. Dell Technologies recently added support for NVIDIA A100 GPGPUs to the PowerEdge R7525 server, which supports up to three PCIe-based dual-width NVIDIA GPGPUs. This blog describes the single-node performance of selected HPC applications with both one- and two-NVIDIA A100 PCIe GPGPUs.

The NVIDIA Ampere A100 accelerator is one of the most advanced accelerators available in the market, supporting two form factors: 

• PCIe version 
• Mezzanine SXM4 version

 The PowerEdge R7525 server supports only the PCIe version of the NVIDIA A100 accelerator. 

The following table compares the NVIDIA A100 GPGPU with the NVIDIA V100S GPGPU: 

The NVIDIA A100 GPGPU brings innovations and features for HPC applications such as the following:

• Multi-Instance GPU (MIG)—The NVIDIA A100 GPGPU can be converted into as many as seven GPU instances, which are fully isolated at the hardware level, each using their own high-bandwidth memory and cores. 
• HBM2—The NVIDIA A100 GPGPU comes with 40 GB of high-bandwidth memory (HBM2) and delivers bandwidth up to 1555 GB/s. Memory bandwidth with the NVIDIA A100 GPGPU is 1.7 times higher than with the previous generation of GPUs. 

Server configuration

The following table shows the PowerEdge R7525 server configuration that we used for this blog:

Benchmark results

The following sections provide our benchmarks results with observations.

High-Performance Linpack benchmark

High Performance Linpack (HPL) is a standard HPC system benchmark. This benchmark measures the compute power of the entire cluster or server. For this study, we used HPL compiled with NVIDIA libraries.

The following figure shows the HPL performance comparison for the PowerEdge R7525 server  with either NVIDIA A100 or NVIDIA V100S GPGPUs:

Figure1: HPL performance on the PowerEdge R7525 server with the NVIDIA A100 GPGPU compared to the NVIDIA V100SGPGPU

The problem size (N) is larger for the NVIDIA A100 GPGPU due to the larger capacity of GPU memory. We adjusted the block size (NB) used with the:

• NVIDIA A100 GPGPU to 288
• NVIDIA V100S GPGPU to 384

The AMD EPYC processors provide options for multiple NUMA combinations.  We found that the best value of 4 NUMA per socket (NPS=4), with NUMA per socket 1 and 2 lower the performance by 10 percent and 5 percent respectively. In a single PowerEdge R7525 node, the NVIDIA A100 GPGPU delivers 12 TF per card using this configuration without an NVLINK bridge. The PowerEdge R7525 server with two NVIDIA A100 GPGPUs delivers 2.3 times higher HPL performance compared to the NVIDIA V100S GPGPU configuration. This performance improvement is credited to the new double-precision Tensor Cores that accelerate FP64 math.

The following figure shows power consumption of the server while running HPL on the NVIDIA A100 GPGPU in a time series. Power consumption was measured with an iDRAC. The server reached 1038 Watts at peak due to a higher GFLOPS number.

Figure2: Power consumption while running HPL

High Performance Conjugate Gradient benchmark

The High Performance Conjugate Gradient (HPCG)  benchmark is based on a conjugate gradient solver, in which the preconditioner is a three-level hierarchical multigrid method using the Gauss-Seidel method. 

As shown in the following figure, HPCG performs at a rate 70 percent higher with the NVIDIA A100 GPGPU due to higher memory bandwidth:  

Figure 3: HPCG performance comparison 

Due to different memory size, the problem size used to obtain the best performance on the NVIDIA A100 GPGPU was 512 x 512 x 288 and on the NVIDIA V100S GPGPU was 256 x 256 x 256. For this blog, we used NUMA per socket (NPS)=4 and we obtained results without an NVLINK bridge. These results show that applications such as HPCG, which fits into GPU memory, can take full advantage of GPU memory and benefit from the higher memory bandwidth of the NVIDIA A100 GPGPU.       

GROMACS

In addition to these two basic HPC benchmarks (HPL and HPCG), we also tested GROMACS, an HPC application. We compiled GROMACS 2020.4 with the CUDA compilers and OPENMPI, as shown in the following table: 

Figure4: GROMACS performance with NVIDIA GPGPUs on the PowerEdge R7525 server

 The GROMACS build included thread MPI (built in with the GROMACS package). All performance numbers were captured from the output “ns/day.” We evaluated multiple MPI ranks, separate PME ranks, and different nstlist values to achieve the best performance. In addition, we used settings with the best environment variables for GROMACS at runtime. Choosing the right combination of variables avoided expensive data transfer and led to significantly better performance for these datasets.

GROMACS performance was based on a comparative analysis between NVIDIA V100S and NVIDIA A100 GPGPUs. Excerpts from our single-node multi-GPU analysis for two datasets showed a performance improvement of approximately 30 percent with the NVIDIA A100 GPGPU. This result is due to improved memory bandwidth of the NVIDIA A100 GPGPU. (For information about how the GROMACS code design enables lower memory transfer overhead, see Developer Blog: Creating Faster Molecular Dynamics Simulations with GROMACS 2020.)

Conclusion

The Dell PowerEdge R7525 server equipped with NVIDIA A100 GPGPUs shows exceptional performance improvements over servers equipped with previous versions of NVIDIA GPGPUs for applications such as HPL, HPCG, and GROMACS. These performance improvements for memory-bound applications such as HPCG and GROMACS can take advantage of higher memory bandwidth available with NVIDIA A100 GPGPUs.