Researcher slams proposal for model-specific inference chips, warns vague definitions could pause non-AI computing
BlancheMinerva · x · 2026-11-02
Blanche Minerva explains a proposal to replace GPUs with chips physically hardwired for specific models (e.g., a chip that can only run DeepSeek 7 13B inference). She calls the idea "complete nonsense": the proposal never clearly defines "training-capable" chips, and she struggles to imagine a definition that wouldn't cause massive collateral pauses for AI-unrelated computing.
Related event: Debate flares over purpose-built inference chips replacing GPUs(3 posts)→
More from Infra
- Prime Mega MoE extends DeepSeek's kernel for training, up to 34% faster end-to-end MoE training — cfregly · 2026-11-02
- AI datacenters priced at ~$50B per gigawatt, but build speed is the metric that matters — JOBhakdi · 2026-11-02
- Reddit asks what happened to 'too cheap to meter' as OpenAI quietly cuts quotas — rolyataylor2 · 2026-11-02
- $1T+ US Mega-Project Pipeline: Texas Eyed as the Future Hub for AI Intelligence Deployment — NinaDSchick · 2026-11-02
- One RTX 3090 Ti runs a 552B DeepSeek MoE at 35 tok/s — full local LLM guide — uniVocity · 2026-11-02
- Musk says robotics needs more DRAM than Earth produces, yet Micron trades at 4x forward PE — sudoraohacker · 2026-11-02