Skip to content
MindPattern
Wire
Briefings
Search
Subscribe
← The Wire
Source trail
Mike Ayles / Hacker News
Public MindPattern findings, entities, and graph evidence that cite this source.
Findings
1
All-time hits
1
High value
0
Last seen
2026-08-11
Related findings
2026-08-11 / NEWS
A 3.16M-Parameter LLM Hits 21,300 tok/s on a $250 FPGA by Never Touching DDR
An August 10 writeup demonstrates a 3.16M-parameter INT4 transformer — about 1.5 MB of weights, trained on TinyStories — running entirely in on-chip SRAM on a $250 Xilinx Kria KV260 board, with zero DDR in the token-generation loop. Measured peak aggregate throughput is 59,965 tok/s at 200 MHz (N=16, bit-exact, three runs of three); the live demo sustains ~21,300 tok/s under 2,000 concurrent connections. For comparison the same board's ARM A53 cores manage 11 tok/s and a laptop RTX 3050 Ti manages 719 — the entire result comes from trading the 20 GB/s shared DDR bottleneck for hundreds of GB/s of on-chip bandwidth, which is the argument for small-model-on-cheap-silicon edge inference in one number.
Open latest cited source
Wire
Briefings
Search
Subscribe