Summary

As we enter the New Year, the market for AI solutions across numerous industries continues to grow. Specifically, UBS predicts a jump from $2.2 billion in 2022 to $255 billion in 2027 [1]. This growth is not limited to large enterprises; GPU support on the new PowerEdge T360 and R360 servers gives businesses of any size the freedom to explore entry AI inferencing use cases, in addition to graphic-heavy workloads. 

We tested both a 3D rendering and AI inferencing workload on a PowerEdge R360 with one NVIDIA A2 GPU[1] to fully showcase the added performance possibilities.

Achieve 5x rendering performance with the NVIDIA A2 GPU

For our first test, we used Blender’s OpenData benchmark. This open-source benchmark measures rendering performance of various 3D scenes on either CPU or GPU. We achieved up to 5x better rendering performance on GPU, compared to the same workload run only on CPU [1]. As a result, customers gain up to 1.70x the performance per every dollar invested on an A2 GPU vs CPU [2].  

Reach max inferencing performance with limited CPU consumption

Part of the motivation behind adding GPU support is the growing demand among SMBs for on-premise, real-time, video and audio processing. Thus, to evaluate AI inferencing performance, we installed NVIDIA’s open-source DeepStream toolkit (version 6.3). DeepStream is primarily used to develop AI vision applications that leverage sensor data and various camera and video streams as input. These applications can be used across various industrial sectors (for example, real-time traffic monitoring systems or retail store aisle footage analysis). With the same PowerEdge R360, we conducted inferencing on 48 streams while utilizing just over 50% of the GPU, and a limited amount of the CPU [3]. Our CPU utilization during testing averaged about 8%.  

The rest of this document provides more details about the testing conducted for these two distinct use cases of a PowerEdge T360 or R360 with GPU support. 

Product Overview

The PowerEdge T360 and R360 are the latest servers to join the PowerEdge family. Both are cost-effective 1-socket servers designed for small to medium businesses with growing compute demands. They can be deployed in the office, the near-edge, or in a typical data analytic environment.

The biggest differentiator between the T360 and R360 is the form factor. The T360 is a tower server that can fit under a desk or even in a storage closet, while maintaining office-friendly acoustics. The R360, on the other hand, is a traditional 1U rack server. Both servers support the newly launched Intel® Xeon® E-series CPUs, 1 NVIDIA A2 GPU, as well as DDR5 memory, NVMe BOSS, PCIe Gen5 I/O ports, and the latest remote management capabilities.

Figure 1. From left to right, PowerEdge T360 and R360

NVIDIA A2 GPU Information

Unlike the analogous prior-generation servers, the recently launched PowerEdge T360 and R360 now support 1 NVIDIA A2 entry GPU. The A2 accelerates media intensive workloads, as well as emerging AI inferencing workloads. It is a single-width GPU stacked with 16GB of GPU memory and 40-60W configurable thermal design power (TDP). Read more about the A2 GPU’s up to 20x inference speedup and features here: A2 Tensor Core GPU | NVIDIA.

Testing Configuration

We conducted benchmarking on one PowerEdge R360 with the configuration in the table below. Similar results can be expected for the PowerEdge T360 with this same configuration. We tested in a Linux Ubuntu Desktop environment, version 20.04.6. 

Table 1. PowerEdge R360 System Configuration

Accelerate 3D Rendering Workloads

Entry GPUs are often used in the media and entertainment industry for 3D modeling and rending. The NVIDIA A2 GPU is a powerful accelerator for these workloads. To highlight the magnitude of the acceleration, we ran the same Blender OpenData benchmark on CPU, and then only on GPU. Blender is a popular open-source 3D modeling software.  

The benchmark evaluates the system’s rendering performance for three different 3D scenes, either on CPU or GPU only. Results, or scores, are reported in sample per minute. We ran the benchmark on CPU (Intel Xeon-E2488) three times, and then on GPU (NVIDIA A2) three times. The results in Table 2 below represent the average score of each of the three trials.   

Results

Compared to the benchmark run only on CPU, we attained up to 5x better rendering performance with the same workload run on the A2 GPU [1]. Although we achieved over 4x better performance for all three 3D scenes, the classroom scene corresponds to the best result and is illustrated in the figure below. 

Figure 2. Rendering performance on CPU only and GPU only

Given this 5x better rendering performance, we calculated the performance per dollar for the cost of CPU compared to the cost of the GPU. For CPU performance, we divided the rendering score by the Dell US list price for the E-2488 CPU. For GPU performance, we divided the rendering score by the Dell US list price for the A2 GPU[2]. When comparing these results, we found customers can gain up to 1.70x the performance per every dollar spent on the GPU compared to the CPU [2]. 

Figure 3. Rendering performance per dollar increase 

Taking the analysis a step further, we also calculated the performance per dollar spent on a CPU compared to cost of both a CPU and GPU. This comparison is relevant for customers who are investing in both an Intel Xeon E-2488 CPU and NVIDIA A2 GPU for their PowerEdge R360/T360. While we calculated the CPU performance score the same way as above, we now divided the GPU rendering score by the Dell US list price for the A2 GPU + E-2488 CPU. When comparing these results, we found customers can gain up to 1.27x the performance per every dollar spent on both GPU and CPU compared to just CPU [2].

In other words, investing in an R360 with a E-2488 CPU and A2 GPU yields a higher return on investment for rendering performance compared to an R360 without an A2 GPU. It is also worth mentioning that the E-2488 CPU is the highest-end, and most expensive, CPU offered for both the T360 and R360. It is reasonable to expect an even higher return on investment for the A2 GPU when compared to the same system with a lower-end CPU. 

The full results and scores are listed in the table below. 

Table 2. Blender benchmark results 

Video Analytics Performance with NVIDIA DeepStream

While 3D rendering may be a more common workload for SMBs investing in entry-GPUs, the same GPU is also a powerful accelerator for entry AI inferencing and video analytic workloads. We used NVIDIA’s DeepStream version 6.3[3] to showcase the PowerEdge R360’s performance when running a sample video analytic application. DeepStream has a variety of sample applications and input streams available for testing. The given configuration files allow you to vary the number of streams for a run of the app which we explain in greater detail below. Input streams can range from photos, video files (with either h.264 or h.265 coding), or even RTSP IP cameras. 

To better illustrate DeepStream’s functionality, consider the images below that were generated from our run of a DeepStream sample app. Instead of using a provided sample video, we used our own stock video of customers entering and leaving a bakery. The AI model in this scenario can identify people, cars, and bicycles. The images below, which are cropped outputs to zoom in on the person at the cash register, show how this vision application correctly identified these two customers with a bounding box and “person” label. 

Figure 4. Cropped output of DeepStream sample app with modified source video

Instead of pre-recorded videos, an RTSP IP camera would theoretically allow a user to stream and analyze live footage of customers in a retail store. Check out this blog from the Dell AI Solutions team for a guide on how to get DeepStream up and running with a 1080p webcam for streaming RTSP output. 

We also tested the DeepStream sample application with one of NVIDIA’s provided videos that shows cars, bicycles, and pedestrians on a busy road. The images below are screenshots of the sample app run with 1, 4, and 30 streams, respectively. In each tile, or stream, the given model places bounding boxes around the identified objects. 

Figure 5. Deepstream sample video output with 1, 4, and 30 streams, respectively 

Performance Testing Procedure

During a run of a sample application, NVIDIA measures performance as the number of frames per second (FPS) processed. An FPS score is displayed for each stream in 5 second intervals. For our testing, we followed the steps in the DeepStream 6.3 performance guide, which lists the appropriate modifications to the configuration file in order to maximize performance. All modifications were made to the source4_1080p_dec_infer-resnet_tracker_sgie_tiled_display_int8.txt configuration file, which is specifically described in the “Data center GPU – A2 section” of the tutorial. Tiled displays like in Figures 4 and 5 above impact performance, so NVIDIA recommends disabling on-screen display/output when evaluating performance. We did the same.

With the same sample video as shown in Figure 5, NVIDIA reports that using an H.264 source, it is possible to host 48 inferencing streams at 30 FPS each. To test this with our PowerEdge R360 and A2 GPU, we followed the benchmarking procedure below: 

1. Modify the sample application configuration file to take in 48 input streams by changing the parameter num-sources to 48, and the batch-size parameter under the streammux section to 48.[4] This is in addition to the other recommended configuration changes described in the guide above. 
2. Let the application run for 10 minutes[5]
3. Record the average FPS for each of the 48 streams at the end of the run
4. Repeat steps 1-3 with 40, 30, 20, 10, 5, and 1 streams. The only modification to the configuration file should be updating the num-sources and batch-size to match the number of streams currently under test. 

Our results are illustrated in the section below. We used iDRAC tools and the nvidia-smi command to capture system  telemetry data every 7 seconds during testing trials as well (i.e. CPU utilization, total power utilization, GPU power draw, and GPU utilization). Each reported utilization statistic (such a GPU utilization) is the average of 100 datapoints collected over the app run period. 

Results

The figure below displays the average FPS (to the nearest whole number) achieved for varying number of streams. As the number of streams tested increases, the FPS per stream decreases. 

Most notably, we achieved NVIDIA’s expected max performance with our PowerEdge R360; We ran 48 streams with an average of 30 FPS each at the end of the 10-minute run period [3]. In general, 30 FPS is an industry-accepted rate for standard video feeds such as live TV. 

Figure 6. DeepStream FPS for varying number of streams

We also captured CPU utilization during our testing. Unsurprisingly, CPU utilization was highest with 48 streams. However, for all number of streams tested, CPU utilization only ranged between about 2-8%. This means most of the system’s CPU was still available for other work while we tested DeepStream. 

Figure 7. CPU utilization for varying number of streams

In terms of power consumption, the figure below shows GPU power draw overlayed on top of total system power utilization. Irrespective to the number of streams, GPU power draw represents only about 25-27% of the total system power utilization. 

Figure 8. System power consumption for varying number of streams

Finally, we captured GPU utilization as number of streams increased. While it varied more so than the other telemetry data, at the max number of streams tested, GPU utilization was about 50%. We achieved these impressive results without driving the GPU to max utilization. 

Figure 9. GPU utilization for varying number of streams

Conclusion

We have just scratched the surface on the performance capabilities of the PowerEdge T360 and R360. Between 3D rendering and entry AI-inferencing workloads; the added A2 GPU allows SMBs to explore compute-intensive use cases from the office to the near-edge. In other words, the R360 and T360 are equipped to scale with businesses as computing demand inevitably, and rapidly, evolves.

While GPU support is a defining feature of the PowerEdge T360 and R360, they also leverage the newly launched Intel^® Xeon^® E-series CPUs, 1.4x faster DDR5 memory, NVMe BOSS, and PCIe Gen5 I/O ports. For more information on these cost-effective, entry-level servers, you can read about their excellent performance across a variety of industry-relevant benchmarks and up to 108% better CPU performance.

References

• Daily: US tech gains set to continue on stronger AI growth | UBS Global
• Blender - Open Data
• DeepStream SDK | NVIDIA Developer | NVIDIA Developer
• A2 Tensor Core GPU | NVIDIA
• Ubuntu 20.04.6 LTS (Focal Fossa)
• Performance — DeepStream 6.3 Release documentation (nvidia.com)
• Understanding FPS Values: The Advanced Guide to Video Frame Rates (dacast.com)
• Battle of the Servers: PowerEdge T360 & R360 outperform prior-gen models across a range of benchmarks | Dell Technologies Info Hub
• Introducing the PowerEdge T360 & R360: Gain up to Double the Performance with Intel® Xeon® E-Series Processors | Dell Technologies Info Hub

Legal Disclosures

[1] Based on November 2023 Dell labs testing subjecting the PowerEdge R360 to Blender OpenData benchmark with 1x NVIDIA A2 GPU and 1x Intel Xeon E-2488 CPU. Actual results will vary. Similar results can be expected on a PowerEdge T360 with the same system configuration.

[2] Based on November 2023 Dell labs testing subjecting the PowerEdge R360 to Blender OpenData benchmark with 1x NVIDIA A2 GPU and 1x Intel Xeon E-2488 CPU. Actual results will vary. Similar results can be expected on a PowerEdge T360 with the same system configuration. Pricing analysis is based on Dell US R360 list prices for both the NVIDIA A2 GPU and Intel Xeon E-2488 processor. Pricing varies by region and is subject to change without notice. Please contact your local sales representative for more information. 

[3] Based on November 2023 Dell labs testing subjecting the PowerEdge R360 with 1x A2 GPU to performance testing of NVIDIA’s DeepStream SDK, version 6.3. We tested the sample application with configuration file named:source4_1080p_dec_infer-resnet_tracker_sgie_tiled_display_int8.txt. The full testing procedure is described in this report. Similar results can be expected with a PowerEdge T360 with the same configuration. Actual results will vary.

Appendix

Dell provides an open-source Reference Toolset for iDRAC9 Telemetry Streaming. With streaming data, you can easily create a Grafana dashboard to visualize and monitor your system’s telemetry in real-time. Tutorials are available with this video and whitepaper. 

The screenshot below is from a Grafana dashboard we created for capturing PowerEdge R360 telemetry. It displays GPU temperature and rotations per minute (RPM) for three fans (we ran the Blender benchmark to demonstrate a spike in GPU temperature). You can also track GPU power consumption and utilization, among many other system metrics. 

Figure 10. Grafana dashboard example