D-RMAP-2 §5a-AMX: the tile arm measured on the AMX tier

[AdaWorldAPI]

Lab probe only, nothing admitted to production. Documentation and board only; no code change.

§5a measured the D-RMAP-2 tile arm on a host with no tile hardware, so its BF16 numbers came from ndarray's F32x16 polyfill. This adds the same measurement on a host where ndarray reports AMX TDPBF16PS.

Method. Same code, fixtures and command as §5a, at main 8b570ed3. CPU flags read before and after the runs (amx_bf16, amx_tile, amx_int8, avx512_bf16, avx512f; 4-vCPU shared VM), unchanged. Five runs of cargo run --release --example tile_gemm_cost, each the example's medians of 7; §5a-AMX reports the median of the five and the range. cargo test --release --test tile_gemm passes on the host, and its BF16 assertions go through the packed entry that dispatches to AMX.

What it shows.

• Exactness holds on AMX: walk counts equal the oracle exactly, and count > 0 equals Boolean composition.
• Walk counts: AMX A·A 58–67 µs against popcount 128–164 µs, about 2.2–2.5×, staging (132–470 µs per map) excluded. By arithmetic from the medians, not measured: staging is repaid after about 1.4–6.4 products on one staged map.
• Boolean composition and triangles: the bit arm still wins, by 6–60× and with 1.9–36 µs against 41–323 µs.
• So §5a's polyfill finding, that the bit arm wins every cell, does not hold on the AMX tier for walk counts. It still holds for existence questions.

Not covered. VDPBF16PS-only hosts, blocked tiles above one focus, the GraphBLAS arm. The row-major entry took 5.6–6.1 ms here against about 19 ms on the §5a host; that path does not use AMX and the difference was not investigated. The example's bf16_bytes column reads 262 144 on this commit; the corrected 384 KiB is in #1625.

Board. STATUS_BOARD D-RMAP-2 row updated; new entry D-RMAP-2-AMX; entries index regenerated (--check passes); supersession index unchanged; append-only gate passes.

---

Generated by Claude Code

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Warning

Billing warning: we have not been able to collect payment for this subscription for more than 72 hours. Please update the payment method or pay any pending invoices in Billing to avoid service interruption.

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-20T09:10:58.290321Z	69e5003	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.