GPUs for prefill, custom silicon for decode: General Compute bets on Positron heterogeneous inference
rohanpaul_ai · x · 2026-10-21
General Compute announced a cluster-scale deployment with Positron: LLM prefill (parallel prompt processing) runs on GPUs, while token-by-token decode is handed to Positron's memory-optimized hardware, sold together as a single endpoint.
The split follows the physics: prefill is a large parallel compute job, while decode is mostly the chip waiting to pull weights and cache from memory—so a single chip must compromise on one. Positron's chips avoid HBM dependency and target decode's memory-bandwidth bottleneck. Hardware goes live Q1 2027.
More from Infra
- NVIDIA Vera Rubin reportedly in full production, demo shows on-chip photonics — LinusEkenstam · 2026-10-21
- NVIDIA Showcases DSX AI Factory Blueprint with GB300 NVL72 Racks — testingcatalog · 2026-10-21
- Data center casualties hit 8 in a month as AI buildout worries grow — AIFlow_ML · 2026-10-21
- Nvidia says Vera Rubin node now built every 5 minutes, unveils $50bn 1GW DSX turnkey AI factory — LinusEkenstam · 2026-10-21
- Jensen Huang unveils Nvidia DSX: $50bn turnkey 1GW AI compute factories — LinusEkenstam · 2026-10-21
- Nextdc Implodes 45% as Angry Residents Cut Power to Melbourne Data Center — AIFlow_ML · 2026-10-21