FIELD NOTE / X
MIG partitions; it does not multiply.
The complete written thought and the evidence behind it. The video edition will follow its public release.
The written argument is here.
This approved X edition is on the journal now. Its video player and original platform link will appear after each public release is verified.
MIG partitions; it does not multiply.
Video caption
MIG partitions; it does not multiply. Free fragments may not fit a larger request. Use MIG when profiles fit and sharing pays. #EricFieldNotes
Full written post / accessibility read
NVIDIA lists up to seven MIG instances on an H200. That is a way to divide one physical device into supported compute and memory profiles. It does not create seven H200s of capacity, or guarantee that the next large job fits the remaining fragments.
Imagine several small tenants occupying different slices. A later request needs one larger profile. The dashboard can show free capacity in aggregate, yet no legal profile layout matches the request without moving or reconfiguring work. Arithmetic free space is not schedulable shape.
Test supported profiles on the actual GPU. Run the intended small and large jobs concurrently, change the mix, and measure admission wait, isolation, performance and the operational cost of any reconfiguration. Compare whole-device scheduling against the same accepted work.
Offer fractional devices only with their exact memory and compute profile and a clear reconfiguration policy. Keep whole-device capacity for jobs that need it. Do this because a slice is a service class, not a smaller label for the same GPU.
#EricFieldNotes
Four-beat scene transcript
1. MIG partitions; it does not multiply.
NVIDIA lists up to seven MIG instances on an H200. That is a way to divide one physical device into supported compute and memory profiles. It does not create seven H200s of capacity, or guarantee that the next large job fits the remaining fragments.
Visual: Seven instances are not seven whole H200s.
2. The wrong shapes strand capacity.
Imagine several small tenants occupying different slices. A later request needs one larger profile. The dashboard can show free capacity in aggregate, yet no legal profile layout matches the request without moving or reconfiguring work. Arithmetic free space is not schedulable shape.
Visual: Free fragments may not fit a larger request.
3. Exercise the actual mix.
Test supported profiles on the actual GPU. Run the intended small and large jobs concurrently, change the mix, and measure admission wait, isolation, performance and the operational cost of any reconfiguration. Compare whole-device scheduling against the same accepted work.
Visual: Small-job gain versus large-job wait and disruption.
4. Partition to the workload.
Offer fractional devices only with their exact memory and compute profile and a clear reconfiguration policy. Keep whole-device capacity for jobs that need it. Do this because a slice is a service class, not a smaller label for the same GPU.
Visual: Use MIG when profiles fit and sharing pays.
Research and claim limits
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.