AVD Pools using AMD CPU/GPU vs. NVIDIA-Intel CPU/.GPU

Hi BrianPitt,

This is a pertinent topic as more organizations look to optimize performance and cost for their Azure Virtual Desktop environments.

Experience with AMD GPU/CPU (NVas V4 Series)

User Experience:

• General Performance: Users have generally reported positive experiences with AMD GPU-based VMs, particularly for workloads that leverage GPU capabilities. The NVas V4 series offers a good balance of price and performance.
• vCPU Allocation: Running 1 user per vCPU on an AMD-based host has been feasible for many workloads. However, the performance can vary depending on the specific applications and their requirements.

GPU Performance:

• Partial vs. Full GPUs: The NVas V4 series VMs with partial GPUs (such as the NV4as V4, NV8as V4, etc.) have shown good performance for applications that do not require full GPU resources. For applications needing intensive GPU processing, the NV32as V4, which uses a whole GPU, is preferable.
• AMD vs. NVIDIA: In comparative tests, AMD GPUs have performed well for a variety of graphical applications. However, NVIDIA GPUs are often preferred for specific tasks such as AI and ML workloads due to their extensive software support and optimized drivers.

CPU Performance:

• AMD CPUs: AMD EPYC processors used in NVas V4 series VMs are competitive with Intel CPUs in many scenarios. Users have found AMD CPUs to provide robust performance for multi-threaded applications and virtual desktop workloads.
• Intel CPUs: Intel-based NVv3 series VMs have historically been favored for single-threaded performance and broader compatibility with enterprise applications. The choice between AMD and Intel often comes down to specific workload requirements and cost considerations.

Issues and Considerations

Partial GPUs:

• Performance Impact: Some users have noticed a slight performance degradation when using VMs with partial GPUs for very GPU-intensive applications. This is expected as partial GPU resources might not meet the demands of high-end graphical applications.
• Optimization: Ensuring that applications are optimized to make efficient use of partial GPU resources can mitigate some of these performance impacts.

CPU Performance:

• AMD vs. Intel: While AMD CPUs perform well, some applications optimized specifically for Intel architecture might run slightly better on Intel-based VMs. It's essential to benchmark specific applications to understand their performance on different CPU architectures.

Practical Steps and Recommendations

1. Benchmarking:

   • Test Workloads: Conduct performance tests for your specific applications on both AMD and NVIDIA-Intel configurations. Benchmarking tools and performance monitoring can provide insights into which configuration offers the best performance for your needs.

2. User Allocation:

   • vCPU per User: Experiment with different user-to-vCPU ratios to find the optimal balance. While 1 user per vCPU is a good starting point, adjustments may be necessary based on the workload.

3. Application Optimization:

   • GPU Utilization: Ensure that applications are optimized for the specific GPU architecture. Some applications might have settings or configurations that can enhance performance on either AMD or NVIDIA GPUs.

4. Cost Considerations:

   • Cost vs. Performance: Evaluate the cost differences between NVas V4 series and NVv3 series VMs. Sometimes, the cost savings with AMD GPUs and CPUs can justify a minor performance trade-off.

Conclusion

Both AMD GPU/CPU and NVIDIA-Intel configurations have their advantages. The choice largely depends on your specific workloads, performance requirements, and budget constraints. Benchmarking your applications on both platforms will provide the best insights into which configuration will deliver the optimal performance for your users.