FIELD NOTE / INSTAGRAM
Partitioning is not multiplication.
The complete written thought and the evidence behind it. The video edition will follow its public release.
The written argument is here.
This approved Instagram edition is on the journal now. Its video player and original platform link will appear after each public release is verified.
Partitioning is not multiplication.
Video caption
Partitioning is not multiplication.
Yesterday's efficient slicing may block today's large job.
Alpha feature gate and driver tags are required.
State capacity, isolation and reconfiguration rules.
#EricFieldNotes
Full written post / accessibility read
MIG can isolate several workloads on one supported GPU. That is valuable when their memory and compute needs match the offered slices. It does not make the original device larger, and not every sum of empty fragments can become the profile a new job needs.
A cluster fills small inference tenants in clean slices. The next customer submits a larger fine-tune. Even if the dashboard has nominal headroom, the offered profile may be wrong. Repacking tenants can cost time or interrupt work, depending on the configuration.
Kubernetes 1.37 introduced alpha Dynamic Resource Allocation compatibility groups, disabled by default. A configured scheduler can reject incompatible MIG or virtual-GPU combinations only when that gate is enabled and the driver supplies group tags. Test those conditions in a disposable cluster, inspect actual admission, then load-test valid profiles. Without that configured path, do not assume the scheduler rejects them.
Quote the actual profile, its memory and compute bounds, expected wait and what happens when demand shifts. Do this because fractional GPU economics only work when the customer job fits the fragment at the time it needs it.
#EricFieldNotes
Four-beat scene transcript
1. Partitioning is not multiplication.
MIG can isolate several workloads on one supported GPU. That is valuable when their memory and compute needs match the offered slices. It does not make the original device larger, and not every sum of empty fragments can become the profile a new job needs.
Visual: A shared device has a finite profile menu.
2. Demand changes the answer.
A cluster fills small inference tenants in clean slices. The next customer submits a larger fine-tune. Even if the dashboard has nominal headroom, the offered profile may be wrong. Repacking tenants can cost time or interrupt work, depending on the configuration.
Visual: Yesterday's efficient slicing may block today's large job.
3. Test configured compatibility early.
Kubernetes 1.37 introduced alpha Dynamic Resource Allocation compatibility groups, disabled by default. A configured scheduler can reject incompatible MIG or virtual-GPU combinations only when that gate is enabled and the driver supplies group tags. Test those conditions in a disposable cluster, inspect actual admission, then load-test valid profiles. Without that configured path, do not assume the scheduler rejects them.
Visual: Alpha feature gate and driver tags are required.
4. Sell a specific slice promise.
Quote the actual profile, its memory and compute bounds, expected wait and what happens when demand shifts. Do this because fractional GPU economics only work when the customer job fits the fragment at the time it needs it.
Visual: State capacity, isolation and reconfiguration rules.
Research and claim limits
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.