GitHub radar
APEX: Open LLM Inference Chip Design for FPGA
Sigmantic AI has published APEX, an open source, bit-exact-verified hardware implementation of a transformer decoder tile for running real LLM models on FPGA chips. It compresses the KV cache in the datapath to maintain constant read throughput during long conversations.
APEX is an open source RTL implementation of a full transformer decoder layer. APEX was created by Sigmantic AI’s autonomous verification agents. At 0.56 tok/s, APEX runs Qwen2.5-0.5B on real FPGAs (that’s 140x speedup). APEX is bit-exact verified against a NumPy golden model before release for all blocks.
Why a vibe-coder should care
APEX achieves constant inference throughput over long conversations through KV cache compression on the device. APEX is fully open source and can be reproduced with a simple git clone — this is basically unheard of in AI hardware!
▌ More finds