FIELD NOTE / INSTAGRAM
Four GPU replicas can still be one GPU.
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.
Four GPU replicas can still be one GPU.
Day 23 · Week 4 editorial group · Instagram · no publication date or time assigned
Video caption
Four GPU replicas can still be one GPU.
Two tenants arrive together and interfere.
Run paired one-tenant and multi-tenant load.
Choose MIG or whole cards when the isolation contract requires them.
#EricFieldNotes
Full written post / accessibility read
NVIDIA time-slicing can advertise several replicas of one physical GPU. That may help small jobs share it, but the replicas do not add memory or proportional compute. Unlike MIG, the time-sliced users do not get memory or fault isolation from the replica boundary.
A quiet single-user benchmark may look fine. Then a second tenant runs a long kernel and the first tenant misses its latency target. Asking for two replicas does not reserve twice the compute. The contract must name the shared-access class honestly.
Use disposable tenants with the same request mix. Measure p ninety-five latency, throughput, memory failures and noisy-neighbor impact. Attribute the outcome with application traces; NVIDIA notes DCGM Exporter container attribution is limited under this time-slicing setup.
Name the fractional mechanism and the guaranteed envelope, if any. Offer a whole-card or MIG class for workloads that need stronger bounds. Do this because multiplying scheduler resources is a placement tactic, not multiplying physical capacity.
#EricFieldNotes
Four-beat scene transcript
1. Four GPU replicas can still be one GPU.
NVIDIA time-slicing can advertise several replicas of one physical GPU. That may help small jobs share it, but the replicas do not add memory or proportional compute. Unlike MIG, the time-sliced users do not get memory or fault isolation from the replica boundary.
Visual: A time-sliced resource is shared access, not extra silicon.
2. The invoice can suggest a guarantee that is absent.
A quiet single-user benchmark may look fine. Then a second tenant runs a long kernel and the first tenant misses its latency target. Asking for two replicas does not reserve twice the compute. The contract must name the shared-access class honestly.
Visual: Two tenants arrive together and interfere.
3. Test the advertised class under neighbors.
Use disposable tenants with the same request mix. Measure p ninety-five latency, throughput, memory failures and noisy-neighbor impact. Attribute the outcome with application traces; NVIDIA notes DCGM Exporter container attribution is limited under this time-slicing setup.
Visual: Run paired one-tenant and multi-tenant load.
4. Sell shared as shared.
Name the fractional mechanism and the guaranteed envelope, if any. Offer a whole-card or MIG class for workloads that need stronger bounds. Do this because multiplying scheduler resources is a placement tactic, not multiplying physical capacity.
Visual: Choose MIG or whole cards when the isolation contract requires them.
Research and claim limits
- NVIDIA GPU time-slicing (S184)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.