用离子材料实现精确浮点计算，突破类脑硬件精度瓶颈
The Native Spiking Microarchitecture: From Iontronic Primitives to Bit-Exact FP8 Arithmetic
- 将噪声神经元当作逻辑单元，设计组合流水线与纠错机制
- 16,129组FP8运算全部与PyTorch结果比特精确一致
- 线性层延迟降为O(log N)，实测提速17倍且抗硬件噪声
2025年诺贝尔化学奖授予金属有机框架（MOFs）研究，加上莫纳什大学王焕庭团队的突破，确立了埃级通道作为后硅时代候选基底，并具备原生积分-发放（IF）动态。然而，如何利用这种随机、模拟的材料实现确定性、比特精确的AI计算（如FP8）仍是一大矛盾。现有类脑方法多依赖近似，无法满足Transformer的精度要求。为跨越‘从随机离子到确定浮点’的鸿沟，我们提出原生脉冲微架构。将噪声神经元视为逻辑基本单元，引入空间组合流水线和粘滞额外修正机制。在全部16,129组FP8运算中验证，结果与PyTorch完全比特精确一致。关键的是，该架构将线性层延迟降至O(log N)，实现17倍加速。物理仿真进一步表明，系统对极端膜泄漏（beta ≈ 0.01）具有强鲁棒性，有效免疫硬件随机性影响。
原文摘要 · Abstract (English)
The 2025 Nobel Prize in Chemistry for Metal-Organic Frameworks (MOFs) and recent breakthroughs by Huanting Wang's team at Monash University establish angstrom-scale channels as promising post-silicon substrates with native integrate-and-fire (IF) dynamics. However, utilizing these stochastic, analog materials for deterministic, bit-exact AI workloads (e.g., FP8) remains a paradox. Existing neuromorphic methods often settle for approximation, failing Transformer precision standards. To traverse the gap "from stochastic ions to deterministic floats," we propose a Native Spiking Microarchitecture. Treating noisy neurons as logic primitives, we introduce a Spatial Combinational Pipeline and a Sticky-Extra Correction mechanism. Validation across all 16,129 FP8 pairs confirms 100% bit-exact alignment with PyTorch. Crucially, our architecture reduces Linear layer latency to O(log N), yielding a 17x speedup. Physical simulations further demonstrate robustness against extreme membrane leakage (beta approx 0.01), effectively immunizing the system against the stochastic nature of the hardware.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。