README: energy row, H1 length-dependence
1 changed file +3 −2
modified
README.md
+3 −2
@@ -17,8 +17,8 @@ Version 0.1 (2026-11-01), **preliminary** working paper; not submitted. Gates A | |||||
| 17 | energy study running at the time of writing, Gate G audit done for the existing logs (see `research/LOG.md`). Nothing in the paper | 17 | energy study running at the time of writing, Gate G audit done for the existing logs (see `research/LOG.md`). Nothing in the paper | ||
| 18 | is reported without a raw log under `benchmarks/raw/`; `artifacts/manifest.json` lists every artifact with its SHA-256. | 18 | is reported without a raw log under `benchmarks/raw/`; `artifacts/manifest.json` lists every artifact with its SHA-256. | ||
| 19 | 19 | ||||
| 20 | −**Verdicts so far (all with bootstrap CIs, `benchmarks/processed/audit.json`):** H1 not supported on OLMoE/laptop NVMe (best predictor ties: | 20 | +**Verdicts so far (all with bootstrap CIs, `benchmarks/processed/audit.json`):** H1 length-dependent on OLMoE/laptop NVMe (64-token runs: best predictor ties, | ||
| 21 | −the install path, not prediction, is the bottleneck); H4 not supported (decode at batch 1 is launch-bound; fusion ±2 %); the remote | 21 | +the install path is the bottleneck; 256-token runs: probe +32 % tok/s); H4 not supported (decode at batch 1 is launch-bound; fusion ±2 %); the remote | ||
| 22 | tier over 1 GbE is a capacity mechanism, not a latency one (−66 tok/s for odd-expert placement); the expert-contiguous storage layout | 22 | tier over 1 GbE is a capacity mechanism, not a latency one (−66 tok/s for odd-expert placement); the expert-contiguous storage layout | ||
| 23 | is the one clear win (cold NVMe 8 → 36 tok/s). | 23 | is the one clear win (cold NVMe 8 → 36 tok/s). | ||
| 24 | 24 | ||||
@@ -39,6 +39,7 @@ constraint: placement decides speed, never which experts run. | |||||
| 39 | | TITAN, 16 slots, cold NVMe + linear probe (recall@8 0.76) | 31.8 | 0.78 | `.../attribution.jsonl` | | 39 | | TITAN, 16 slots, cold NVMe + linear probe (recall@8 0.76) | 31.8 | 0.78 | `.../attribution.jsonl` | | ||
| 40 | | Qwen3-30B-A3B on M3U96a: mlx-lm / TITAN all-resident | 121.7 / 121.6 | — | `benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl` | | 40 | | Qwen3-30B-A3B on M3U96a: mlx-lm / TITAN all-resident | 121.7 / 121.6 | — | `benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl` | | ||
| 41 | | OLMoE two nodes (M3U96a + M3U96b, 1 GbE), 32 slots: local / odd experts remote | 81 / 15 | — | `benchmarks/raw/gateE/two_node_m3u96a_m3u96b.jsonl` | | 41 | | OLMoE two nodes (M3U96a + M3U96b, 1 GbE), 32 slots: local / odd experts remote | 81 / 15 | — | `benchmarks/raw/gateE/two_node_m3u96a_m3u96b.jsonl` | | ||
| 42 | +| Energy, 256-token runs: all-resident 0.101 J/token (47 W); 16 slots cold 0.498 J/token (31.9 tok/s); + probe 0.441 J/token (**42.0 tok/s**, +10.1 [+5.2, +16.9]) | — | 0.63 / 0.84 | `benchmarks/raw/gateG/energy_laptop/` | | ||||
| 42 | 43 | ||||
| 43 | ## Repository layout | 44 | ## Repository layout | ||
| 44 | ``` | 45 | ``` | ||
| 45 | 46 | ||||