Tag
Akhetonics 展示了其全光计算原型——一颗可编程的光子 CPU，目标实现太赫兹级时钟速度，数据全程以光形式处理而无需光电转换。文章介绍了其两比特光学处理器架构、近 30 代芯片迭代以及计划于 2026 年底部署首台商用机器，但也指出了内存、封装、功耗和制造等方面尚未解决的挑战。
UCLA researchers unveiled a hybrid optical-neural processor that detects deepfake videos with ~97.79% accuracy while analyzing 15-18 video streams in parallel, using light propagation to replace energy-hungry digital decoding. The system achieves 99.86% sensitivity and can be deepened with passive diffractive layers for higher accuracy at minimal energy cost.
A talk on moving beyond GPUs and backpropagation: the speaker argues that the co-evolution of hardware and optimizers has led to stagnation in FLOPs per joule, claims that within a decade we will abandon backpropagation, and proposes pairing SPSA zeroth-order optimizers with metamer-based hybrid architectures, while betting on alternative paths such as optical computing and neuromorphic computing.
Y Combinator's Paper Club session explores AI compute beyond the transformer-backprop-GPU stack, covering optical computing with light, neuromorphic chip architectures inspired by the brain, and biological neurons trained with reinforcement learning to play Doom.
AI hardware startup Volantis announced an $88M Series A to use optical interconnects for chips with vastly increased memory capacity and bandwidth, enabling ultra-fast inference of up to 10,000 tokens per second for models over 10T parameters — a potential unlock for coding agents.
The author explores an alternative approach to AI perception using wave superposition instead of neural network weights, seeking insights from researchers in related fields.
This repository presents a proof of concept for a hardware-native optical timing-frozen control plane engine that uses fluid dynamics modeling in GPU registers to reduce jitter in optical data centers for distributed AI architectures.
Otters++ is a novel optical spiking Transformer that leverages time-to-first-spike coding and physical hardware decay to achieve energy-efficient inference, achieving 84.17% on GLUE while maintaining a clear energy advantage over prior spiking Transformer baselines.
This paper proposes DxPTA, a novel design space exploration methodology for efficient HW/SW co-design of photonic transformer accelerators that meet area, power, energy, and latency constraints. It achieves up to 15.2x faster searching time than exhaustive approaches, enabling efficient PTA designs for diverse transformer models.