Four DGX Sparks, one week of tuning: 26.85 to 56.71 tok/s
AIFlow_ML · x · 2026-11-21
A week of tuning four DGX Spark boxes lifted inference throughput from 26.85 to 56.71 tok/s.
Key wins:
- depth-3 drafter for speculative decoding
- FP8 attention (+12%)
- faster path for small all-reduces (+16%)
- raising the sequence cap cut the slowest first token at 8 concurrent users from 64s to 1s
Takeaway: tune the four boxes you have before buying a fifth.
More from coding & agent
- Japanese Devs Say Personal 'Software Garages' Are Ready to Be Shown — 4310sy · 2026-11-21
- Lessons from shipping a no-auth read-only travel MCP server through Claude's directory review — Wooden_Body_7891 · 2026-11-21
- call.md: open-source desktop app turns meetings into live agent loops, 2.1k GitHub stars — tom_doerr · 2026-11-21
- Dev swaps GLM 5.3 Flash for Thinking Machines' Inkling 2 Small in Windows observability app — michellechen · 2026-11-21
- Dev fully automates Honkai Star Rail with Claude Code + GPT Sol on a Pixel 10 in ~5 days, ~$300 — Ok_Chipmunk_8734 · 2026-11-21
- I replaced automation chains with 31 tiny single-job agent apps — NexelythCollective · 2026-11-21