Home > Servers > Rack and Tower Servers > AMD > White Papers > Workload-Based DDR5 Memory Guidance for Next-Generation PowerEdge Servers > Computer vision and 3D rendering
To measure the performance on different memory configurations for computer vision and 3D rendering workloads, we used Blender, which is an open-source 3D creation and modeling software project. The Blender benchmark includes subtests such as BMW27, Classroom, Barbershop, Fishy Cat, and Pabellon Barcelona, which test Blender's cycles performance with various sample files. We ran all the subtests in the CPU-only mode.
The most common performance unit used in Blender benchmarking is the time taken to complete a specific rendering or simulation task. The time taken is usually measured in seconds or minutes and represents the total duration required to finish the given workload. In this context, a lower time-taken value indicates better performance, meaning the task was completed more quickly.
The following figure shows the performance of Blender (BMW27) across R7615 and R7625 servers with 4th Gen AMD EPYC 9654 and 9654P processors:
Figure 13. Performance trend for BMW27 with different DIMM configurations and memory capacity for 4th Gen AMD EPYC 9654 and 9654P processor-based PowerEdge R7625 and R7615 with default BIOS settings
The time taken to perform the workload in PowerEdge R7625 was almost 80 percent lower than that in PowerEdge R7615. We looked at performance per watt and cost effectiveness for the 6-DIMMs-per-socket configurations. The performance is similar for all memory configurations in PowerEdge R7625. The following figures show the performance of Blender BMW27 across the R7615 and R7625 servers with 4th Gen AMD EPYC 9654 and 9654P processors:
Figure 14. Performance per 1,000 W trend for Blender BMW27 in 6-DIMMs-per-socket, nearly balanced configuration and memory capacity for 4th Gen AMD EPYC 9654 processor-based PowerEdge 7625 with default BIOS settings
Figure 15. Performance per $100,000 trend for Blender BMW27 in 6-DIMMs-per-socket, nearly balanced configuration and memory capacity for 4th Gen AMD EPYC 9654 processor-based PowerEdge 7625 with default BIOS settings
The performance per watt in the 6-DIMMs-per-socket configuration with 16 GB DIMMs was 34.5 percent better than the 6-DIMMs-per-socket, balanced 32 GB configuration, and it was 85 percent better than the 6-DIMMs-per-socket, nearly balanced, 64 GB configuration. Similarly, the performance per dollar in the 6-DIMMs-per-socket, 16 GB configuration was 9.5 percent better than the 6-DIMMs-per-socket, balanced 32 GB configuration and 29.2 percent better than the 6-DIMMs-per-socket, nearly balanced 64 GB configuration.
We find it to be a CPU-capacity-bound workload. Based on test results, we recommend PowerEdge R7625 featuring AMD EPYC 9654P with a configuration of 6 DIMMs per CPU. Any memory size can be used, as long as the model can fit into the memory and there are enough slots for growth and future scalability.