diff options
Diffstat (limited to 'docs')
| -rw-r--r-- | docs/hardware/HW_RESEARCH_FINDINGS.md | 13 |
1 files changed, 9 insertions, 4 deletions
diff --git a/docs/hardware/HW_RESEARCH_FINDINGS.md b/docs/hardware/HW_RESEARCH_FINDINGS.md index d380b39..e6b77b9 100644 --- a/docs/hardware/HW_RESEARCH_FINDINGS.md +++ b/docs/hardware/HW_RESEARCH_FINDINGS.md @@ -242,7 +242,12 @@ SGD/Lion: 1.0× memory-cells, ~0 ADC/step, trivial periphery ← analog-n Adam(fact): 2.0× memory-cells, ~0 ADC/step, +16C AGC channels ← analog-compatible Muon: 2.0× memory-cells, +12C² ADC conversions/step + digital NS sidecar ← mixed-signal heavy ``` -Recommendation for the chip narrative: **Lion (or SGDM) as the on-chip rule; factored-Adam as the -premium tier; Muon explicitly priced out** (its value is in the GPU/simulation era, where it is -nearly free). Pending verification: Tiki-Taka transfer constants, AGC channel specs, Loihi-trace -leakage numbers — the deferred optimizer-state deep-research (rerun when API budget recovers). +**CORRECTION (2026-07-11, project empirical record): Lion/SGD are RULED OUT for EP training** — +tested and failed (looped line: abl_lion2 diverged at step 600, val 20.1 / res 0.34; earlier tests +concur). Per-parameter adaptivity (Adam-class) is empirically LOAD-BEARING for EP gradients here. +⟹ Revised chip narrative: **factored-Adam (AGC/Tiki-Taka route) is not the premium tier — it is +the MANDATORY baseline**. The +2× analog memory and ~16C AGC channels per block are the price of +admission for analog EP training, and the OmniOpt "Lion is cheap"路线 does not apply to us. +Muon stays priced out of analog (GPU/simulation-era tool). This RAISES the priority of the deferred +optimizer-state research (Tiki-Taka transfer constants, AGC channel specs, trace leakage — rerun +when API budget recovers): the analog-Adam engineering is now on the critical path of the chip story. |