Exstill 5nm co-designed processor lowers token energy

arxiv.org1 min read

Researchers introduce Exstill, a co-designed single-chip processor and model architecture optimized to lower the energy and latency floor for language model token generation. Holding weights and conversation key-value caches entirely on-die, a 5 nm pre-silicon design brings Qwen3-1.7B token generation close to its physical arithmetic floor and achieves massive speedups and energy reductions compared to standard hardware like the NVIDIA B300. Additionally, the authors present Polyglot-DC variants specialized for tasks such as speculative drafting, multilingual guarding, and financial reasoning. Open article →