SemiAnalysis — TileRT compiles the whole model into one persistent GPU kernel, claiming 1.9x decode interactivity on the same Blackwell chips

X post

SemiAnalysis — TileRT compiles the whole model into one persistent GPU kernel, claiming 1.9x decode interactivity on the same Blackwell chips

Software squeezing 2x more interactive inference out of the same NVIDIA silicon cuts at the core pitch of the dedicated inference-ASIC vendors (Groq, Cerebras, SambaNova) — and raises effective utilization per deployed GPU, which feeds the power-per-rack math the AI-power lane tracks.

Open on X
SemiAnalysis✓ @SemiAnalysis_
Aug 11

ALERT🚨 TILERT from@TileRT_AII TO BOOST DECODE INTERACTIVITY BY 1.9X AT THE SAME PER-TOKEN COST ON THE SAME NVIDIA BLACKWELL GPUs. What does this mean for Groq/Cerebras/SambaNova? 👇️ 1/5pic.x.com/lTBJA1rgO0gO0

Image attached to the post

Why it's worth your time

Software squeezing 2x more interactive inference out of the same NVIDIA silicon cuts at the core pitch of the dedicated inference-ASIC vendors (Groq, Cerebras, SambaNova) — and raises effective utilization per deployed GPU, which feeds the power-per-rack math the AI-power lane tracks.

7 events
DateTypeEvent
Aug 8 tweetIPO Newsroom — NVIDIA is putting up to $3B into the Blackstone-backed firm that owns the land under Stargate Jul 28 ArticleFifteen of twenty broke together Jul 26 redditu/rdh2dmd — Samsung, SK Group seal $950B in AI deals with Nvidia and Broadcom Aug 28 ArticleThe Layers That Feed AI Compute All Trade Below Their Averages. Compute Doesn't. Jul 22 pickAlphabet Q2: $44.9B quarterly capex, $180-190B FY guide, 2027 'significantly increases' Sep 16 tweetSemiAnalysis — Rubin NVL72 benches at 7x the performance per watt of GB300, not the 3x claimed Jun 22 ArticleWeek in Review — The Rotation Down-Stack

Tweets, videos, filings, articles, and source pages tied to this read.

1
X@SemiAnalysis_ tweet