Question. ExpertPlex does bounded tile-level preemption inside a
persistent kernel with Hopper primitives: thread-block clusters, DSMEM,
and TMA multicast. These do not exist below sm_90. What preemption
granularity, cost, and safety are possible without them?
Answer. Bounded preemption works on Ampere. Claim N1 holds on both
pre-registered axes, and the decomposition repeats on sm_80. One
scope limit is itself a measured result: checkpoint cost follows
occupancy. The full results are in RESULTS.md.
AI tools (Claude) assisted with code and documentation. The developer
directed, reviewed, and verified all work. No measured number comes
from a model: every result comes from the instruments in src/, gated
by pre-registered thresholds (CLAIM.md), float64
references, sanitizer runs, and reproduction on a second GPU.
Status. All measurement phases complete. The paper draft is written. Phase 6 (real model) is optional.
| Measurement | Result |
|---|---|
| Tile calibration | K = 32 to 128 gives 6.7 to 17.8 us per tile at 1200 MHz |
| Baseline, no preemption | Urgent task waits 957 us (median) at full residency |
| Yield at tile boundaries | 54 to 69 us median latency across builds, at most 3.4% cost |
| Scaling surface, 24 cells | N1 holds. Propagation stays at 1 to 3 us at all block counts |
| Mid-pipeline yield safety | 0 of 28,859 buffer-conflicting yields corrupt. 17.4 us median latency, in the Hopper band |
| sm_80 anchor (A100) | The decomposition transfers. Two memory-headroom predictions failed informatively |
| Reserved capacity | Cheap but cliffs in waves and binds to resource envelopes. Cooperative holds the better points |
| Scheduler oracle | ~185M task executions, each exactly once. Zero violations |
| Tensor-core pipeline | Safety and oracle transfer whole. Per-stage checkpoint cost is 14% at 1 block/SM; tile-boundary cost is 3% |
Hardware. NVIDIA A10 (GA102, sm_86, 24 GB, 72 SMs), SM clock
locked and verified under load (1200 MHz; one part holds 1050 MHz). One
A100-SXM4-40GB (sm_80, 1410 MHz) as the second-architecture anchor. A
second A10 reproduced the primary results within 2%.
| File | Content |
|---|---|
RESULTS.md |
All results, limits, and conclusions |
CLAIM.md |
Claims N1 and A-N1 with pre-registered thresholds. Never edited after registration |
BACKGROUND.md |
Definitions and the Hopper primitives in question |
FINDINGS.md |
Literature facts with citations |
ROADMAP.md |
Phases, kill criteria, and measurement rules |
NOTEBOOK.md |
Dated run log: prediction, result, faults found |
papers/ |
Reading index (PDFs not committed) |
paper/ |
The paper source and figures |
src/ |
Five binaries, one per measurement. Each file header says how to run it |
scripts/ |
One script per measurement, plus lock_clocks.sh and the table generators |
results/summary/ |
Committed figures and summary CSVs. Every quoted number traces to one of these files |
Report what fails. Two prior repos (https://github.com/AyaanFaisal21/TransformerOp), (https://github.com/AyaanFaisal21/cornfield) set this standard. This repo keeps it: 18 faults found, all recorded in the NOTEBOOK, none deleted. Registered predictions that failed are reported as findings, not removed.