Understanding the Nodes panel

Under the Cluster Strategy option, you can locate the Nodes panel. This panel lists all the MIG capable nodes.

Each node lists its available GPUs, and for each GPU the currently associated slices.

As an example, the following illustration contains a cluster that has two MIG capables nodes:
  • ecs-sxm-0631.ecs.mc.comops.rocks
  • ecs-sxm-0633.ecs.mc.comops.rocks
These two nodes are identical, have the exact same GPUs: 8 x NVIDIA H100 80GB HBM3.

Additional information about these slices and how these should be read:

  • NVIDIA H100 80GB HBM3 has capacity: 80gb memory and 7 compute units.
  • Slices like: 1 * 3g.40gb + 2 * 1g.20gb implies that we split that GPU to 3 slices overall:
    • 1 * 3g.40gb: This single slice would use 40GB memory and 3 compute units
    • 2 * 1g.20gb: This yields 2 slices of the same size, each would have 20GB memory and 1 compute unit
    • This uses up the whole GPU memory: 40gb + 2 * 20gb = 80gb
    • However, not all the compute units are used: 3 + 2 * 1 = 5. This GPU has 7 compute units, so 2 would be left unused.
    • Because these slices exhaust the GPU capacity, there is no option to add more slices to this GPU.
  • There is a special slice, that is visible on the sceenshot for ecs-sxm-0633.ecs.mc.comops.rocks GPU 0: 1g.10gb+me. This is a special slice and only 1 could be used on a GPU, you cannot have 2 slices of the same. The +me means to include media engines such as: NVDEC, JPEG, OFA.