EVENT DOSSIER
The researchers proposed a time-domain Softmax circuit based on an RC temperature-adjustable circuit, which is implemented in GlobalFoundries’ 22-nm FDSOI process. This architecture operates directly on the voltage generated in memory, without the need for intermediate analog-to-digital conversion. Exponential weights are generated through shared falling斜坡 and RC attenuation reference values, and normalization is performed. The 128-element implementation occupies an area of 9453.42 μm², with each element occupying 70.2 μm²; the evaluation delay at a power consumption of 13.44 mW is 242.97 ns, and the energy consumption per element is 25.5 pJ. After Monte Carlo analysis and extraction of parasitic parameters, the root mean square error (RMSE) for the simultaneous evaluation of 128 elements is 24.46 mV. The circuit characteristics have been integrated into the MemTorch hardware-aware Transformer model.
研究人员提出了一种基于 RC 可调温度的时域类软最大（Softmax）电路，该电路运行于 GlobalFoundries 22-nm FDSOI 工艺中。该架构直接在存内计算（CIM）生成的电压上操作，无需中间模数转换，通过共享下降斜坡和 RC 衰减参考值生成指数权重并进行归一化。128 元素实现占用面积 9453.42$\mu\mathrm{m}^{2}$，单个元素占 70.2$\mu\mathrm{m}^{2}$；在 13.44 mW 功耗下评估延迟为 242.97 ns，单元素能耗为 25.5 pJ。经蒙特卡洛分析及寄生参数提取验证，128 元素同时评估的均方根误差（RMSE）为 24.46 mV。该电路特性已集成至 MemTorch 硬件感知 Transformer 模型中，其验证损失在理想软最大基线…