Neural network runs on FPGA with no CPU, OS, or software — pure Verilog logic
blaizedsouza · x · 2026-09-05
Zakriya Paracha built an MNIST handwritten-digit classifier that exists entirely as hardware inside an FPGA: a 28×28-input feedforward net with 25,888 weights, quantized from float to 8-bit integers, shrinking the model from 98 KB to 25 KB with only 0.18pp accuracy loss. Every multiply, memory access, and decision runs as dedicated Verilog-synthesized circuits — no CPU, no OS, no software — enabling ultra-fast inference.
More from Infra
- Extropic teases Z1T-0.1 as it races from GPT-2 toward GPT-3-level models — beffjezos · 2026-09-05
- Extropic founder claims 1000X more efficient AI inference could be worth trillions — beffjezos · 2026-09-05
- Podcast Jokes About Scouting a 20GW Datacenter Site in Tahoe — chrisalbon · 2026-09-05
- Datacenters vs Shale Oil: Investors Wrong to Dismiss Selling Compute as a Commodity — abhiadesai · 2026-09-05
- Omnigent's Four-Line Defense Against Surprise LLM Bills: Routing, Budgets, Visibility — matei_zaharia · 2026-09-05
- GPT-6 Astra reportedly pretrained on 100,000+ GPUs at Stargate in Abilene, Texas — NinaDSchick · 2026-09-05