Signal Neuromorphic Spike-Based Large Language Model (NSLLM) & Edge Neuromorphic Chips 2026

Summary

A spike-based large language model, described in a paper published 24 December 2025 in National Science Review (DOI 10.1093/nsr/nwaf551), achieved 19.8 times the energy efficiency, 21.3 times the memory efficiency and 2.2 times the inference throughput of an A800 GPU baseline. Implemented on a VCK190 FPGA using a MatMul-free computing core, the billion-parameter model drew 13.849 watts of dynamic power while running at 161.8 tokens per second. On vision tasks neuromorphic hardware delivered around 1,000 inferences per joule against 10 to 100 for GPUs, with the human brain's sub-20-watt consumption cited as a reference point. For specific workloads energy savings reached three to five orders of magnitude, and the system performed competitively on common-sense reasoning, reading comprehension, world knowledge and mathematics tasks versus mainstream models of comparable scale. Full market entry is expected from 2027.