DiffusionGemma Hits ~6k tok/s on H200 Estimate, Making Uno Paper's Plot 'Extremely Suspicious'
bodonoghue85 · x · 2026-09-06
bodonoghue85 notes that a benchmark plot runs DiffusionGemma on a single H100, and estimates it could top 6,000 tok/s on an H200.
Based on this, he calls the corresponding plot in the Uno paper 'extremely suspicious,' implying its baseline numbers may not hold up. A substantive challenge to diffusion-LLM performance benchmarking.
More from Models
- GPT-6 Astra Draws a Unicorn for $0.11, 15x Cheaper Than GPT-5.4 Pro — zakelfassi · 2026-09-06
- GPT-6 Astra hits 97.2% on short-doc extraction SOTA but only 31.7% on long docs — llama_index · 2026-09-06
- HF Engineer: Circuit Board Design Is My Personal Turing Test for AI — and Models Are Catching Up — hugs · 2026-09-06
- Google Astra timeouts still burn Plus-plan credits, user reports — sharkbaitlol · 2026-09-06
- Astra Minecraft test shows high reasoning one-shots a whole house via command block — hankberger · 2026-09-06
- Astra voice mode a huge step up, works well as orchestrator — jdjohnson · 2026-09-06