Benoit Jacob bjacob

Functionality implications of keeping Sampled as the only MOI engine

Scope and evidence

Source review of shared/rocjitsu/sanitizers, revision 75881fde3e181682f8bcddca58bc8031106757e1. All source paths below are relative to rocm-systems. No new GPU trials or test runs were performed. Fault specifications are executable expectations, not freshly measured detection rates. Status ledgers summarize previous accepted evidence; neither substitutes for rerunning the exact binary/configuration after changes.

Benchmark update: the gfx950 campaign documents were read after pulling revision 3ddf9dee36. They now contain the completed physical MI350X results. The implementation and test discussion below remains the original source review at 75881fde3e; its line references refer to that revision. This update does not constitute a full re-review of the intervening implementation changes or a rerun of the fault corpus. The benchmark evidence strengthens the performance motivation for Sampled but does not estab

ConSan `gfx1201` benchmark status

For each mode, Startup is the total latency through the first synchronized run and its selected evidence checkpoints, including instrumentation, loading, binding, and warm-up; Run is the absolute second-run latency followed by its ratio to the matching uninstrumented second run.

Workload	Uninstrumented Startup	Uninstrumented Run	SuperCollider Startup	SuperCollider Run	RecordReplay Startup	RecordReplay Run	Sampled Startup	Sampled Run	InlineShadow Startup	InlineShadow Run
PyTorch synthetic dense prefill (32-token prompt)	0.937 s	0.00209 s (1×)	132 s	0.00332 s (1.59×)	182 s	0.805 s (386×)	163 s	0.0092 s (4.41×)	159 s	0.431 s (207×)

ConSan Synchronization Identities and the SGEMM Scalability Gap

What is a synchronization identity?

ConSan assigns a stable textual identity to each synchronization event found while analyzing a GPU code object. A representative identity has this shape:

fnv1a64:<code-object-fingerprint>
|kernel=

Current matrix

Every workload retained in the active matrix is simultaneously green. Final audit artifact consan-validation-gfx1250-final-audit-149 reruns every one of the 40 cells at committed tip 9acc4dd9b0 with one hook identity. Cell text reports progress within blue; successful process launch or focused unit coverage alone does not promote a cell.

ConSan gfx1250 status

This is the gfx1250 workload × instrumentation evidence ledger. It follows the acceptance standard of STATUS_RDNA4.md, but inherits no coverage denominator, machine-code identity, fault expectation, timing, provenance, or green cell from another architecture.

The executable authority is consan_validation.py, with the experiment contract described by VALIDATION.md. Porting work

RocJitsu vs. FFM gfx1250 benchmark

Measured on 2026-07-19 using the same gfx1250 workload binaries and the workspace TheRock ROCm runtime. Each reported row passed its independent correctness oracle. FFM used its default worker configuration; forcing eight workers was separately found to cause severe slowdowns and is not represented in the table.

Test	RocJitsu	FFM	Comparison

Priority workload	SuperCollider	Record/Replay	Sampled	Inline Shadow
P0 Qwen3-0.6B prefill	🟩 20/20; clean + exact-fault trap; overhead: 1.0x	🟩 20/20; clean + qualified miss; overhead: 1.0x	🟩 20/20; clean + measured statistical detection; repeated-dispatch metadata hardening tracked; overhead: 1.0x	🟩 20/20; clean + attributed diagnostic; overhead: 1.0x
P1 Sharktank TP1 prefill	🟩 Clean 352/352; exact pair removal is a schedule-masked qualified miss; instability marker retained; overhead: 1.1x	🟩 Clean 352/352 + 92/92 barriers; exact pair removal corrupts oracle, qualified replay miss; overhead: 1.7x	🟩 Clean 352/352; exact pair removal is a schedule-masked qualified miss; four unpatched supported barriers are decode-only; overhead: 1.0x	🟩 Clean 352/352 + 92/92; exact pair removal emits an attributed diagnostic; overhead: 4.4x
P1 Sharktank TP1 decode/combined	🟩 Clean 352/352; exact pair removal is a schedule-masked qualified miss; c

Priority workload	SuperCollider	Record/Replay	Sampled	Inline Shadow
P0 Qwen3-0.6B prefill	🟩 20/20; clean + exact-fault trap; overhead: 1.0x	🟩 20/20; clean + qualified miss; overhead: 1.0x	🟩 20/20; clean + measured statistical detection; repeated-dispatch metadata hardening tracked; overhead: 1.0x	🟩 20/20; clean + attributed diagnostic; overhead: 1.0x
P1 Sharktank TP1 prefill	🟩 Clean 352/352; exact pair removal is a schedule-masked qualified miss; instability marker retained; overhead: 1.1x	🟩 Clean 352/352 + 92/92 barriers; exact pair removal corrupts oracle, qualified replay miss; overhead: 1.7x	🟩 Clean 352/352; exact pair removal is a schedule-masked qualified miss; four unpatched supported barriers are decode-only; overhead: 1.0x	🟩 Clean 352/352 + 92/92; exact pair removal emits an attributed diagnostic; overhead: 4.4x
P1 Sharktank TP1 decode/combined	🟨 Clean decode + combined; 352/352; instability marker	🟨 Clean decode +

	How /goal was part of the problem

	/goal installs a hook that runs every time the main thread ends a turn. A checker reads the conversation and
	decides whether the goal is met. If it is not, the session is nudged to continue. That produced two costs.

	- A per-turn tax. The cost is the number of turn ends times the context size. That night there were 299 turn
	ends with the goal active, at about 500k tokens each. About a quarter of the bill is cache reads and writes
	that appear in the session's cost record but in no transcript. The goal checks fit that volume. The docs say
	the checker runs on Haiku, but the cost record shows almost no Haiku usage. A 500k conversation also does
	not fit Haiku's window. So the checks were probably billed at Fable rates, but I have not proven it.


	› Why does extreme register pressure pass more easily on emulation than on physical GPU ?


	• Context compacted

	• The key distinction is that Rocjitsu models architectural register values, while the physical GPU also enforces the kernel’s declared
	register allocation and real instruction-dependency behavior. I’m checking the exact Rocjitsu paths now so I can separate what we know
	from the current leading hypothesis.