live eventEidos is moving itself to 90%-cheaper silicon without losing intelligence. Watch.follow on LinkedIn →
mission: 90% local authorship — waiting for first eventhow?
← charts + scores

llama3.2:1b on A6000, A100, H100 — 2026-04-18

2026-04-18T00:32:34.000Z 2026-04-18T00:33:13.251Z
tweet →

On 2026-04-18, llama3.2:1b ran across A6000, A100, H100. H100 finished first at 589.0 tok/s, 22× faster than A6000. The per-million-token cost: $1.17 (H100) vs $3.57 (A6000) — the surprise: the headline GPU H100 was also the cheapest per-million tokens, coming in 3.0× less than A6000.

Podium

H100
1st
thunder-h100
peak tok/s
589.0
avg tok/s
589.0
$ / 1M tok
$1.17
A100
2nd
thunder-a100
peak tok/s
49.1
avg tok/s
49.1
$ / 1M tok
$4.41
A6000
3rd
thunder-a6000
peak tok/s
27.2
avg tok/s
27.2
$ / 1M tok
$3.57

silicon telegraphH100 just posted 193 tok/s on llama3.1:8b — 13.7× faster than the A100 lane (1428h ago).

llama3.2:1b on A6000, A100, H100 — 2026-04-18 — Eidos Live · Crucible