diff options
Diffstat (limited to 'docs/outreach/OUTREACH_TARGETS.md')
| -rw-r--r-- | docs/outreach/OUTREACH_TARGETS.md | 14 |
1 files changed, 13 insertions, 1 deletions
diff --git a/docs/outreach/OUTREACH_TARGETS.md b/docs/outreach/OUTREACH_TARGETS.md index 28934ef..0d1b621 100644 --- a/docs/outreach/OUTREACH_TARGETS.md +++ b/docs/outreach/OUTREACH_TARGETS.md @@ -68,7 +68,19 @@ plan are attached. Best, Yuren -### Draft A (wave-1b) — Shanbhag group · To: Soonha Hwang (soonhah2@), Mihir Kavishwar (mihirvk2@) · cc: Shanbhag +### ⛔ Draft A SUPERSEDED 2026-07-13 — DO NOT SEND AS-IS (M-scale down-select §6) +The published record answers the die-mode questions NO: the 65nm DIMA trainer's outputs are 3 +comparators + a 6b test-ADC pair (JSSC-2018, author PDF), and the 28nm ESSERC chip is a **Digital +IMC** by the group's own Nov-2025 slides. Pitching "your CIM as our EP substrate" would show we +hadn't read their papers. The INVERTED ask (amended wording from the audit, use verbatim or close): +"We've read the JSSC-2018 die docs — the 512x256 bank's outputs are comparator decisions plus the +6b test ADC pair, and the 28nm DiT is digital IMC per your Nov-2025 deck. Two questions: does ANY +die variant in the group's inventory expose per-column analog outputs before aggregation, and would +you co-design a per-column two-phase contrast-readout macro on a future run?" Plus: harness/gateware +collaboration + digital-IMC baseline are their real near-term value; the T64 gateware ports +unchanged to any future analog macro. Rewrite Draft A around this before sending. + +### Draft A (wave-1b, STALE — see banner above) — Shanbhag group · To: Soonha Hwang (soonhah2@), Mihir Kavishwar (mihirvk2@) · cc: Shanbhag Subject: Backprop-free transformer training — GPU-scale results and a staged path to CIM Hi Soonha and Mihir, |