CU masking works on gfx90a including under graph replay, and buys isolation not concurrency

[Andrei-Dr]

Phase 0 for the prefill/decode co-scheduling idea, measured on GPU0 (idle box) 2026-08-09. docs/58 + the probe source in benchmarks/cu_mask_probe.hip. Runs in under a minute, needs no model.

Four of five hardware assumptions hold. The fifth is the interesting one.

gate	result
hipExtStreamGetCUMask read-back	IDENTICAL for 16 / 52 / 104-bit masks
placement	16 bits -> exactly 16 physical CUs; 52 -> 52; 104 -> 104, read from HW_ID per workgroup, not from hipSuccess
disjoint masks	0 CU overlap between bits 0-51 and 52-103
HIP graph replay	mask HELD — a graph captured on a 16-CU stream replays on 16 CUs; the same graph on an unmasked stream spreads to 104
concurrency	disjoint masks 1.99x vs serial — and two ordinary unmasked streams also 1.99x

The graph row was the one I expected to fail (undocumented; the mask binds to a dedicated HSA queue at stream creation, and hipGraphLaunch could plausibly route around it). It does not fail, which matters because vLLM decode is graph-replayed.

The last row is why this is not a green light. HIP already overlaps independent streams here, so CU masking buys isolation, not concurrency — a real but strictly smaller property than the framing invites. And the probe kernel is pure-ALU spin, so it says nothing about HBM contention, which is where the interference actually lives given round 62 has decode at 10% of achievable bandwidth.

Two traps documented for whoever builds on this:

• the flat mask is de-interleaved as SE = bit % 8 on 8 SE x 13 CU, so 13 contiguous bits span all eight shader engines, not one. Any "give decode N adjacent CUs" sweep is measuring a different partition than it thinks.
• a masked stream takes a dedicated HSA queue against a 32-queue / 8-per-ACE budget, and the mask is creation-time only — no setter.

Framed against the incumbent rather than against nothing: this has to beat docs/52's already-adopted DP=2 (1.118x aggregate, 0.888x TPOT), and it contends for the board power budget docs/53 shows is already binding on prefill — a kill criterion the original framing omitted.

Ledger entry added. No production change.