Skip to content

dev: X31: one-dimensional double on Zen 5, Granite Rapids and Axion (cloud) - #20

Merged
jtravs merged 4 commits into
mainfrom
exp/x31-cloud-1d
Nov 16, 2026
Merged

jtravs merged 4 commits into
mainfrom
exp/x31-cloud-1d

Conversation

@jtravs

@jtravs jtravs commented Nov 16, 2026

Copy link
Copy Markdown
Owner

Measurement only, no code change. Experiment file dev/experiments/X31-cloud-1d.md, PLAN index row, LESSONS 2.9, ledger section.

Setup. Google Cloud Spot, us-central1-a: c4d-standard-16 (AMD EPYC 9B45, Zen 5), c4-standard-16 with minimum platform Granite Rapids (Xeon 6985P-C), c4a-standard-16 (Axion, Neoverse V2); x86 with one thread per core. Ubuntu 24.04, GCC 14.2.0, FFTN Release. Tracking tier, one thread, fbench pinned to one core with taskset, 40 rounds each, no round under load. Arms: N = main 4ebcee5 (before X30b), O = ad6acbe (before the push), F = FFTW 3.3.11 built on each machine with its SIMD and FMA codelets, M = oneMKL 2026.2.1 wheel (Intel only).

Runs. 20261116-115045-ed43 (Zen 5, band ±1.24 %), 20261116-115322-2631 (Granite Rapids, ±1.00 %), 20261116-114945-a77e (Axion, ±1.00 %). Stores are kept locally under bench-out/runs, not committed (as for X21).

FFTN ÷ FFTW over the 46 one-dimensional double cases (before the push in brackets): Zen 5 0.80 (1.05), Granite Rapids 0.88 (1.12), Axion 0.74 (0.86); Granite Rapids against MKL 0.85 (1.09). Per-group tables and the worst ten per machine are in the experiment file.

Divergences for lead (d): Axion beats the Mac's ratios on long rows (0.75 against 1.19), in place, real rows and planes, mostly because FFTW is much slower there than on the Mac; Axion 8 points 1.24 (Mac 0.96). Granite Rapids repeats the Xeon's long mixed rows, worse (100000 2.16, 46080 1.60), and in place 65536 is 1.50. Zen 5 diverges in no group; it does not share the EPYC's 65537 loss.

Cost USD 0.75 of the USD 20 cap; every VM, the network, subnet and firewall rule deleted by recorded ID and the teardown verified by listing.

🤖 Generated with Claude Code

jtravs and others added 4 commits November 16, 2026 11:45
…pids, Axion)

Written before the runs: question, arms (main, ad6acbe, FFTW, MKL on Intel),
method and the divergence rule.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… Axion

Tracking tier, one thread pinned, 40 rounds each, no round under load, on
Google Cloud Spot VMs (c4d-standard-16, c4-standard-16 on Granite Rapids,
c4a-standard-16), GCC 14.2.0, arms main 4ebcee5, ad6acbe, FFTW 3.3.11 with
its SIMD, and MKL on Intel. FFTN / FFTW over the 46 one-dimensional double
cases: Zen 5 0.80, Granite Rapids 0.88 (MKL 0.85), Axion 0.74. Runs
20261116-115045-ed43, -115322-2631, -114945-a77e (bands 1.24, 1.00,
1.00 %). Divergences for lead (d) in the experiment file; LESSONS 2.9;
ledger USD 0.75, teardown verified. No code change, no gate stage run.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@jtravs
jtravs merged commit 6a74375 into main Nov 16, 2026
12 checks passed
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