This work presents a tunable-temperature analog softmax circuit in GlobalFoundries 22-nm fully depleted silicon-on-insulator (FDSOI) technology that operates directly on CIM-generated score voltages without intermediate analog-to-digital conversion.
Abstract
Softmax is a key operation in Transformer attention, but its exponentiation and normalization add significant overhead in compute-in-memory (CIM) accelerators, especially when analog attention scores must first be converted to the digital domain. This work presents a tunable-temperature analog softmax circuit in GlobalFoundries 22-nm fully depleted silicon-on-insulator (FDSOI) technology that operates directly on CIM-generated score voltages without intermediate analog-to-digital conversion. Each input score is converted into a time-domain event using a shared falling ramp. The corresponding comparator transition samples an RC-decaying reference to generate an exponential weight, which is then processed by an in-circuit normalization stage. In contrast to analog softmax circuits that rely on transistor weak-inversion behavior for exponentiation, the proposed architecture controls the softmax response through the ramp slope and RC time constant, enabling programmable effective temperature. The 128-element architecture is evaluated using transistor-level and post-layout extracted simulations, including multi-level input vectors, capacitance variation and mismatch, process and temperature variation, monte carlo analysis, and shared-interconnect parasitics. The complete 128-element implementation occupies 9453.42~$\mu\mathrm{m}^{2}$ including the shared global ramp circuitry, while each replicated softmax element occupies 70.2~$\mu\mathrm{m}^{2}$. The circuit achieves a 242.97-ns evaluation latency at 13.44~mW total power, corresponding to 25.5~pJ per output element. The simultaneous 128-element evaluation achieves an RMSE of 24.46~mV relative to the ideal softmax response. The extracted circuit characteristics are further incorporated into a MemTorch-based hardware-aware Transformer model, where the proposed softmax achieves a validation loss within 2.5\% of the ideal-softmax baseline.
Charge-CIM addresses the bottleneck in ACiM accelerators by using switched-capacitor charge redistribution as a unified computing and conversion substrate, reducing both standalone converter overhead and intermediate ADC invocations.
Zihao Xuan, Ye-Wen Li, Jia Chen et al.· 0 citations
Modern edge devices increasingly require real-time adaptation to their environment without relying on cloud-based updates, which can introduce latency and security risks. To meet these demands, memory-augmented neural networks (MANNs) have gained traction for enabling adaptive on-device learning. Hardware implementatio...
Changhoon Joe, K. Byun, Min-Seung Kang et al.· npj Unconventional Computing· 0 citations
This work presents a schematic-level digital near-memory computing architecture based on a 1-Transistor-3-Resistor (1T3R) bit-slicing scheme for 3-bit signed weight storage and indicates that the proposed architecture can maintain functional classification capability under 3-bit weight and 2-bit input constraints.
Zeyuan Hou, Xiao-Meng Wang, Yang Yi· Journal of Electronics and E...· 0 citations
This work introduces MITRA, a reconfigurable magnetic tunnel junction-based in-memory architecture that leverages stochastic computing (SC) to implement a broad class of transcendental and nonlinear functions directly within memory.
Farzad Razi, M. Moghadam, M. Najafi et al.· International Symposium on L...· 0 citations
SRAM-based computing-in-memory (SRAM-CIM) alleviates the memory-wall bottleneck of the von Neumann architecture, enabling energy-efficient AI edge computing. Current-domain CIM schemes suffer from degraded linearity at low supply voltages, whereas time-domain CIM schemes are highly sensitive to process, voltage, and te...
Xiao-Bo Gong, Bin Qiang, Zi-Li Jiang et al.· IEEE Transactions on Circuit...· 0 citations
The architecture thereby combines significance-aware time encoding, channel-scalable readout, and hardware-aware adaptation, and combines significance-aware time encoding, channel-scalable readout, and hardware-aware adaptation.
Ao-Ming Zhan, Ye Zhao, Yu-Mei Zhou et al.· Applied Sciences· 0 citations
Related blog posts
Microsoft Research Blog· microsoft.comNov 16, 2026
AI coding agents are increasingly help developers fix bugs and build features largely on their own, but every interaction with their models consumes tokens. In each session, an agent sends instructions, tool descriptions, and session history to its model. Most efforts to reduce this cost focus on individual tasks. Yet developers return to the same […] The post TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient appeared first on Microsoft Research.
MIT News · Artificial Intelligence· news.mit.eduNov 13, 2026