GLM 5.3 Flash on 4x 8 GB/Confirmation 6
From CMP Wiki
More actions
| Confirms | GLM 5.3 Flash on 4x 8 GB |
|---|---|
| Result | Reproduced |
| By | @darkmaniac7 |
| Date | 2026-10-19 |
| Cards | 8 GB |
| Numbers | Recipe 1.8.0 as published, MiaAI protocol, Gen2 x8, no P2P, 74 SMs, 180 W: PP4 one user 201 / 149 / 97 tok/s (structured / code / prose), 664 / 548 / 477 at eight users, 32k cold prefill 6,834-6,977 tok/s; TP4 361 / 284 / 199, 697 / 587 / 529 at eight users, prefill 1,633-1,753. On an 8-card box at Gen2 x16 with P2P (abliterated W4A16 checkpoint, own bench with 1,024-token answers, code / structured / prose): TP4 252 / 235 / 139, 623 / 617 / 421 at eight users, prefill 1.7k. |
PP4 reproduces the published numbers closely; TP4 lands at about 0.75x without x16 links and P2P, because the all-reduce runs through host memory. Draft acceptance matched the published regime. Vision and video worked. Source: @darkmaniac7, 2026-10-19 (lab notes)
Source: @darkmaniac7, 2026-10-28 (lab notes)