Recommended NVIDIA GPUs for NVIDIA RTX vWS
Table 2 lists the hardware specification for the most recent generation NVIDIA data center GPUs recommended for NVIDIA RTX Virtual Workstation.
L40S 2 |
A16 3 |
L4 |
A40 |
A10 |
|
---|---|---|---|---|---|
GPUs/ Board (Architecture) | Ada Lovelace | Ampere | Ada Lovelace | Ampere | Ampere |
Memory Size | 48GB GDDR6 with ECC | 4x 16 GB GDDR6 with ECC | 24GB GDDR6 with ECC | 48GB GDDR6 with ECC | 24GB GDDR6 with ECC |
vGPU Profiles | 1GB, 2GB, 3GB, 4GB, 6GB, 8GB, 12GB, 16GB, 24GB, 48GB | 1GB, 2GB, 4GB, 8GB, 16GB | 1GB, 2GB, 3GB, 4GB, 6GB, 8GB, 12GB, 24GB | 1GB, 2GB, 3GB, 4GB, 6GB, 8GB, 12GB, 16GB, 24GB, 48GB | 1GB, 2GB, 3GB, 4GB, 6GB, 8GB, 12GB, 24GB |
Form Factor | PCIe Full Height Full Length adapter 4.4” (H) x 10.5” (L), dual slot | PCIe 4.0 Dual Slot Full Length Full Slot (FHFL) | PCIe low-profile, 1-slot | PCIe 4.0 Dual Slot Full Length Full Slot (FHFL) | PCIe 4.0 Single Slot Full Length Full Slot (FHFL) |
Power | 350W | 250W | 72W | 300W | 140W |
Thermal | Passive | Passive | Passive | Passive | Passive |
Use Case | Accelerates deep learning and machine learning training and inference. Along with Light to High-end 3D design and creative workflows. Flexibly runs mixed workloads for both virtual workstations and compute workloads | Entry level Virtual Workstations Upgrade path for M10 | End-to-end acceleration for the next generation of AI-enabled applications from gen AI, LLM inference, small-model training and fine-tuning to 3D graphics, rendering, and video applications Upgrade path for T4 | Light to High-end 3D design and creative workflows. Flexibly runs mixed workloads for both virtual workstations and compute workloads | Ideal for mainstream professional visualization applications running on high-performance mid-range virtual workstations. |
It is essential to resize your environment when switching from Maxwell GPUs to newer GPUs like Pascal, Turing, and Ampere GPUs. For example, the NVIDIA T4 leverages ECC memory which is enabled by default. When enabled, ECC has a 1/15 overhead cost due to the need to use extra VRAM to store the ECC bits themselves; therefore, the amount of frame buffer usable by vGPU is reduced. Additional information for each hypervisor can be found in the respective NVIDIA documentation accessible here.
With the Ada and Ampere architectures, increased frame buffer (FB) requirements are crucial to consider. It is not recommended to use 1 or 2 GB profiles due to their limitations in meeting modern workload demands. For GPUs like L40S or A40, using small profiles can quickly lead to channel limitations. Therefore, opting for larger FB profiles is essential for optimal performance.
Important points to consider:
Modern Workload Demands: Applications such as high-resolution graphics, AI, and data-intensive tasks require significant GPU memory. Small FB profiles (1-2 GB) are inadequate for these applications, leading to frequent memory overflows and degraded performance.
Channel Limitations: GPUs have a limited number of channels. Smaller FB profiles quickly exhaust these channels, preventing efficient parallel processing and leading to an application error.
Performance Optimization: Larger FB profiles provide the necessary memory bandwidth and capacity to handle complex workloads efficiently. With larger profiles, you can run fewer vGPUs simultaneously, but they receive more channels, making it less likely to encounter channel limitations and ensuring smooth and consistent performance.
Scalability: Investing in larger FB profiles not only meets current demands but also provides a buffer for future workload increases, reducing the need for frequent upgrades.
For more details on GPU channel calculations, see Understanding GPU Channels.
Virtual GPU Type |
Frame Buffer (GB) |
Maximum vGPUs per GPU in Equal-Size Mode |
Maximum vGPUs per GPU in Mixed-Size Mode |
Use Case |
---|---|---|---|---|
A40-4Q | 4 | 12 | 8 | Virtual Workstations (vWS) |
A40-8Q | 8 | 6 | 4 | Virtual Workstations (vWS) |
L40S-8Q | 8 | 6 | 4 | Virtual Workstations (vWS) |
L40S-16Q | 16 | 3 | 2 | Virtual Workstations (vWS) |
For more details on Equal-Size and Mixed-Size modes, see vGPU Profiles.
For detailed configurations and additional guidelines, please refer to the NVIDIA vGPU User Guide.
For the L40S and A40 GPUs, 1GB profiles are not recommended. Instead, larger profile sizes are advised to fully utilize the capabilities of these GPUs. This is particularly important because these GPUs are designed to handle high-performance workloads that demand more substantial resources.
NVIDIA A16 is recommended only for entry level virtual workstations with light weight users. A minimum 8GB (8Q) profile is recommended when deploying NVIDIA RTX Virtual Workstations with A16.