Jia-Jun Lu, Xiao-Yang Zeng, Chi-Xiao Chen, et al.
IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · 2026-11-01
SpikeLet combines hierarchical spike aggregation, cross-layer and cross-timestep parallelism, and communication-aware mapping for multichiplet spiking neural networks. On a 4-by-4-chiplet prototype, the authors report a 74.2% reduction in transmission latency in their evaluation. The design addresses the communication and scheduling costs that become important when a spiking network grows beyond one chip's capacity.
SpikeLet面向多chiplet脉冲神经网络,结合分层脉冲聚合、跨层与跨时间步并行,以及通信感知映射。在4×4个chiplet的原型上,作者报告其评估中的传输延迟降低74.2%。该设计针对脉冲网络规模超过单芯片容量后逐渐突出的通信与调度开销。
S. Addy, Eduardo Calle-Ortiz, Prashant Shenoy, et al.
arXiv:2611.03735 · 2026-11-01
HyVox assigns speech embeddings to a neuromorphic processor and speaker matching, unknown-speaker rejection, and enrollment to a microcontroller. Its backbone compresses 6.5-fold to 29,000 synapses, while a 32-neuron embedding extractor passes only spike counts between the processors. This partition lets new speakers enter through gallery-prototype updates and concentrates sparse event-driven computation on the neuromorphic chip.
HyVox将语音嵌入计算交给类脑处理器,将说话人匹配、未知说话人拒识及注册交给微控制器。其骨干网络压缩6.5倍至29000个突触,32神经元嵌入提取器仅通过脉冲计数与另一处理器交换信息。这种划分允许通过更新说话人库中的原型注册新用户,并将稀疏事件驱动计算集中到类脑芯片上。