Yesterday I wrote a blog about how to configure the M60 / M6 NVIDIA GPU boards for use for “graphics” mode or “compute” mode and how they were designed for different use cases such as VDI accelerated graphics and HPC (high-performance compute) respectively.
This has led to some questions as to whether “graphics” supports compute and CUDA like technologies. The answer is yes! The graphics mode simply ensures that limitations imposed by third-party components such as hypervisor, OS and graphics drivers do not affect the use of the GPU.
This means in graphics mode CUDA, OpenCL and similar compute is available to the graphics stack and software features built on top. Software features available in the GRID 2.0 offerings such as GPU-passthrough and vGPU GPU-sharing have access to the full CUDA capabilities of the Tesla M60 hardware.
Both the GRID K1/2 and the Maxwell GPUs such as M60 fully support CUDA and OpenCL.
Currently the vGPU feature has only enabled CUDA and OpenCL in the Mx8Q profiles on cards like the M60 where a vGPU is in fact a full physical GPU, i.e. an equivalent configuration to GPU pass-through. This has benefits for monitoring the GPU from the hypervisor which is not possible with GPU-passthrough.
CUDA/OpenCL are not available on the older GRID K1/K2 cards even where a vGPU represents a whole GPU. I wrote a blog pertaining to the K1/K2 cards here, which explains the technical limitations on vGPU profiles that only represent part of the card.
Virtualisation stacks and OSs
Many OSs and virtualization stacks (VMware, Citrix etc) or even 3D applications may impose limitations on how and whether a GPU is used. But fundamentally the hardware you buy with a M60 does support CUDA, OpenCL and compute whether it is used in “compute” or “graphics” mode. For VDI accelerated graphics though “graphics” mode is our recommended configuration for the M60 / M6 cards, as per yesterday’s blog.