Skip to content

Add MrTukTuk RTX 5090 community runs (PCIe 5.0 x16, 800 W) - #471

Open
bovali wants to merge 1 commit into
jesdga95:mainfrom
bovali:community/mrtuktuk-5090
Open

bovali wants to merge 1 commit into
jesdga95:mainfrom
bovali:community/mrtuktuk-5090

Conversation

@bovali

@bovali bovali commented Nov 19, 2026

Copy link
Copy Markdown

Two runs on the same rig, both re-benched at the card's full PCIe 5.0 x16 width (800 W power limit), basalt-bench --community (community-1), engine d3b7574, Windows, Ryzen 7 9800X3D 8c, 64 GiB:

  • Qwen3.8-27B NVFP4 — decode 575.0, prefill 16K 18,709.7 / 64K 13,322.1 tok/s. The 27B sits fully in 32 GiB VRAM, so the x8 → x16 gain is small (~+1.5 % decode), as expected for a VRAM-resident pack.
  • Qwen3.8-Flash-Next IQ3_XXS — decode 535.3, prefill 16K 10,392.0 / 64K 11,658.6 tok/s. 81 GiB MoE streaming experts over PCIe; the wider link is the whole point of this row.

Both rows pass tools/check_benchmarks.py.

@giveen

giveen commented Nov 19, 2026

Copy link
Copy Markdown
Contributor

800W, i would be scared of that dying.

This branch has not been deployed

No deployments
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.

2 participants