FIELD NOTE / LINKEDIN
Use MIG where workload shape fits.
The complete written thought and the evidence behind it. The video edition will follow its public release.
The written argument is here.
This approved LinkedIn edition is on the journal now. Its video player and original platform link will appear after each public release is verified.
Use MIG where workload shape fits.
Video caption
Use MIG where workload shape fits.
Fragmentation is a product risk.
Run profile, occupancy and turnover experiments.
My rule: Price fit, isolation and transition costs.
#EricFieldNotes
Full written post / accessibility read
Fractional GPUs can make inference sharing economical. But 'one seventh of an H200' is not a complete service specification. NVIDIA documents defined MIG profiles with finite memory and compute. The customer needs to know which profile they receive and how that profile performs under their workload.
A tenant mix optimized for small instances may strand capacity for larger jobs. Reconfiguration can change admission and disruption behavior. The provider's utilization score may improve while a high-value job waits for a legal whole-device or large-profile allocation.
Test the actual application memory and throughput on each offered profile. Simulate the customer mix, arrival bursts, profile changes and whole-device demand. Measure accepted work, tenant isolation, wait time and recovery cost. Let admission reject incompatible combinations before execution.
Offer exact MIG profiles with a visible queue and reconfiguration policy. Maintain whole-device options where needed. Do this because a fleet of partitions is a flexible product only when its shapes match the jobs customers actually submit.
#EricFieldNotes
Four-beat scene transcript
1. Use MIG where workload shape fits.
Fractional GPUs can make inference sharing economical. But 'one seventh of an H200' is not a complete service specification. NVIDIA documents defined MIG profiles with finite memory and compute. The customer needs to know which profile they receive and how that profile performs under their workload.
Visual: The SKU must include the partition profile.
2. The fleet may be full of unusable pieces.
A tenant mix optimized for small instances may strand capacity for larger jobs. Reconfiguration can change admission and disruption behavior. The provider's utilization score may improve while a high-value job waits for a legal whole-device or large-profile allocation.
Visual: Fragmentation is a product risk.
3. Model the mix before pricing it.
Test the actual application memory and throughput on each offered profile. Simulate the customer mix, arrival bursts, profile changes and whole-device demand. Measure accepted work, tenant isolation, wait time and recovery cost. Let admission reject incompatible combinations before execution.
Visual: Run profile, occupancy and turnover experiments.
4. Contract for the slice you can deliver.
Offer exact MIG profiles with a visible queue and reconfiguration policy. Maintain whole-device options where needed. Do this because a fleet of partitions is a flexible product only when its shapes match the jobs customers actually submit.
Visual: Price fit, isolation and transition costs.
Research and claim limits
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.