#2462 2026-10-14: Radeon VII 16 GB + Radeon Pro VII / MI50 16 GB (gfx906, mixed pair), Ryzen 7 3700X, 62 GB RAM

Descrição

Partial benchmark data point (full COMMUNITY_BENCHMARKS PR to follow after the 32K/128K sweep + needle checks).

Source build of v0.1.42 (`61b3fb5`) on the gfx906 opt-in path, HIP 7.14 from the community image
`mixa3607/rocm-gfx906:7.14-complete`, Coder IQ1_M native pack, layer split 27 across both cards,
`--kv int8 --kv-resident 32768`, `--max-context 32768`.

This closes the Radeon VII half of the `docs/AMD_HIP.md` "Not done: … MI60 and Radeon VII (the same gfx906 ISA;
not run)" line. The pair is a Radeon VII (`0x66af`) with a Radeon Pro VII / Instinct MI50 (`0x66a1`) — same ISA,
mixed cards. HIP names both "AMD Radeon VII", so the existing MI50 report's hardware line is ambiguous in the
same way.

`STRATA_STAGE_TRIM=1` puts **all 12,265 experts resident** (6,889 + 5,376 slots), which reproduces the MI50
report's "all 12,288 experts in VRAM" on a mixed pair.

Two configuration results that differ from the published gfx906 report (which used `--spec 4 --spec-min-p 0.5`):

- **`--spec-min-p 0.7` instead of 0.5** — two independent scans, n=7 per value: 4K decode 68.6 → 71.5 (+4.2%),
  16K 65.5 → 67.6 (+3.2%). 0.5 is the worst value tested; the minimum spans 0.2–0.9 and 0.7 ties with 0.8.
- **`--spec 2` instead of 4** — interleaved A/B, 18 samples: 4K decode median 73.40 → 77.50 (+5.6%, mean +6.9%),
  and the spread narrows from sd 2.2 to 0.73. Draft acceptance 77.9% → 94.3%. Note `--spec N` expands to
  `mtp_max_t = N` with verify window `min(N+2, 8)` (suffix drafting defaults to 3), so the window goes 6 → 4.

Also measured on this pair: `STRATA_SPLIT_MTP_BATCH=1` gives prefill +2.0% / +2.3% / +2.3% at 1.5K / 3K / 4K
(sd < 0.5 over 3 restarts), and `STRATA_STAGE_TRIM=1` gives decode +3.2% (4K) / +5.9% (16K) with `route tail skip`
misses falling from ~1,400 to ~19.

### Not done / limits
- Prompt sweep is 4K and 16K only; no 32K or 128K, no needle check, no TTFT table. `--max-context 32768` bounds
  the current config.
- Measurements are controlled short-prompt bench (256 generated tokens, greedy, one request at a time), not
  production mixed traffic — do not compare them against the mixed-traffic numbers in the MI50 report.
- Model is on local NVMe. Storage is a first-order variable on this box: the same model on JuiceFS over a
  gigabit LAN measured 43.4 tok/s where local NVMe measured ~69.4 tok/s.