SymbolicLight V2 introduces a hybrid neuromorphic architecture for language inference, combining sparse event computation with continuous-state processing. The 194M-parameter model replaces standard softmax with local attention and uses graded signed events to optimize performance. Benchmarks on an Alveo U50C FPGA show a 27.6% reduction in gross card energy per token, dropping to 0.04407 J, while increasing decode throughput from 474.6 to 643.2 tokens/s.
- Hybrid neuromorphic design merges sparse spikes with continuous states for efficient inference.
- Softmax-free local attention and graded signed events reduce computational overhead.
- FPGA implementation achieves 27.6% energy savings per generated token vs prior version.
- Decode throughput on U50C FPGA rises to 643.2 tokens/s with optimized KV loading.
