agpt-2b across machines⚓︎

One config, three systems: agpt-2b, tp=1, bs=1, seq 2048, 20 iters. The point is a like-for-like number, so the config string is identical everywhere and appears verbatim in each runs.csv.

Results⚓︎

Per device, one A100 does ~1.95x a PVC tile, at 1.87x the MFU. Aggregate favours Sunspot only because that node has 12 tiles to Perlmutter's 4.

The Sunspot row was re-measured on 2026-08-25 (job 12473794) because the original had no MFU or TFLOPS — it predates the current export and its run directory was gone. Throughput reproduced to within 1.22% of the 2026-08-11 figure (61,965 vs 61,217 tok/s) across two weeks and a different commit, which is the main reason to trust either number.

One caveat the ratio hides: a PVC tile is not an A100. Aurora/Sunspot expose 12 tiles per node (2 per GPU, 6 GPUs), so "per-device" means per-tile there and per-card here. Per node is the fairer comparison for capacity planning; per-device is the efficiency one.

There is no matched Aurora row. 80b_benchmarks.md is a different model entirely (80B, tp>=2, 5 steps) and is not comparable.

Reproducing⚓︎

Perlmutter:

sbatch -N1 --export=ALL,BS=1,SEQ=2048,COMPILE=no,VENV=.venv-213 \
    experiments/perlmutter/amsc_bench.sbatch

Sunspot/Aurora, same config through PBS:

python3 -m ezpz.launch --np 12 -- python3 -m ezpz.examples.fsdp_tp \
    --model agpt-2b --tp 1 --dataset random \
    --train-iters 20 --batch-size 1 --seq-len 2048 \
    --outdir "$PWD/outputs/agpt2b-xmachine"

ezpz export-amsc <run-dir> --config agpt-2b/bs1/seq2048/tp1