Skip to content

docs: N1 diagnosis — fresh quiet-host ratios, counters, causes ruled out - #85

Merged
tannevaled merged 1 commit into
mainfrom
docs/n1-diagnosis
Nov 1, 2026
Merged

tannevaled merged 1 commit into
mainfrom
docs/n1-diagnosis

Conversation

@tannevaled

Copy link
Copy Markdown
Contributor

Docs only. This follows up the open defect recorded in #83 (products slow down on a busy N1 host).

  • Fresh quiet-host ratios: measured back to back with NumPy on the same pinned core (load 2): float64 1.09/1.07/1.16, float32 1.08/1.00/1.07 at 64²/256²/1024². So the defect is the sensitivity to a busy host, not the quiet-host speed.
  • Where the gap is: hardware counters (perf stat, perf record -e ll_cache_miss_rd) put it in the micro-kernel's memory behaviour. Against OpenBLAS: 4.5× L1D refills, 1.6× L2, 1.7× last-level misses, mostly in gemmUA4x8D.
  • What was ruled out: six candidate causes, each by A/B with counters. They are a regression, reading A in place, block sizes, writing C, software prefetch, and OpenBLAS's whole structure at once.
  • What remains: a rewrite of the N1 kernel after dgemm_kernel_8x4.S.
  • Fix: a sentence left in the wrong paragraph by an earlier edit is corrected.

No code change, so no release.

🤖 Generated with Claude Code

…le out

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@tannevaled
tannevaled merged commit ca032bc into main Nov 1, 2026
12 checks passed
@tannevaled
tannevaled deleted the docs/n1-diagnosis branch November 1, 2026 09:44
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant