About A100 card and difference between 7 MIG and 4 GPU


We are currently using NVIDA M10 card, where each card has 4 GPUs of 8 GB each. In order to use applications which requires CUDA cores, we have to provide 8 GB GPU to a virtual machine, in which the application is installed. If we assign 1 GB or 2 GB or 4 GB GPU, the CUDA is not detected by the application. So it seems CUDA functionality can be leveraged, only when an entire 8 GB GPU is assigned. Hence, we can create 4 such virtual machines which is CUDA enabled.

We are now trying to upgrade to a better performing GPU. M10 is a high density GPU and we are looking for high performance card for applications like Blender. I searched on internet and found A100 can be a good option. However, I have below questions.

  1. How many CUDA Cores A100 card has?
  2. A100 has 4 GPUs and 7 MIG, what is the difference?
  3. What is MIG instances?
  4. Is it same as GPU slicing?
  5. Can this be done on Hyper-v?

Also if you have any suggestions for any other high performance card to be used on physical server, where it can be assigned to virtual machines, please let me know.


This statement does not seem correct. As far as I can tell, the M10 comprises four GPUs, each with 8 GB of memory attached. The A100 on the other hand is a single GPU that can be partitioned in up to 7 multi-GPU instances (MIG). I have not used an A100 or MIG, but from the documentation it seems like you could partition a A100 with 40 GB of on-board memory into five MIG instances with 8 GB each. The partitions do not have to be of the same size, but in your use case that seems to make the most sense.

Maybe someone with more A100 experience will come along and address the rest of these questions.

The problem with M10 is that, although 8 GB card can be partitioned to 1 GB, 2 GB and 4 GB, but when we assign the partitioned GPU to a virtual machine, CUDA is not detected by most application. For CUDA to be detected, we have to provide the 8 GB card to a VM.
With MIG, can I assign smaller partitioned GPU to a VM, with CUDA still enabled/available?
Can use Microsoft Hyper-V for virtualization and use any of Ampere series card and be able to partition it (I know this is not possible with M10 card as drive is available only for VMware and Citrix Xen server)?