The charts in this section show the key conclusions and takeaways from the test results.
AI duration and AUC accuracy
The following figure shows that in the dual workload GPU-enabled workloads, AI and VDI together perform significantly better for model training times and AUC accuracy compared to CPU-only configurations.
This chart depicts both the duration (left Y axis) and AUC accuracy (right Y axis) for four test cases. AI training and validation without GPU for both single and dual workload test cases are used as baseline measurements. GPU enhancement significantly reduces duration and increases accuracy. The AI workload is 28 times faster with GPUs and the dual AI and VDI medical workload is 10 times faster with GPUs. At the same time, in both instances model accuracy is higher with GPU acceleration.
Dual AI and VDI workloads details
The following figures show the model AUC accuracy and Login VSI base response times respectively.
These dual workloads can co-exist and are not impacted by “noisy neighbor” type issues. Both the AI and VDI metrics show that the dual test case scenario performed well for model accuracy and VDI user experience response times.
Workload vGPU and vCPU utilization profiles
The following figures show several views of resource utilization with a combined view of CPU and GPU resource utilization. This can provide insight into the impact of switching workloads between VDI and AI in a single use case operation when compared with running dual workloads in parallel, and how either case can impact overall utilization.
Duration of workloads
The following figure shows how the duration of workload execution is significantly reduced by applying GPU resources.