🧲 China's Hanxu Shows Inference Architecture Based on Magnetic Memory
On August 31, 2026, Hanxu Technology (寒序科技), reportedly founded in 2023 at the Center for Applied Magnetism of Peking University, unveiled uHBM and uLPU architectures for large model inference: a Compute-Memory Die crystal combines non-volatile magnetoresistive memory (MRAM) and matrix-vector computations, so model weights are constantly resident in memory, and only activation vectors enter and exit the crystal. The claimed metrics are on-chip read bandwidth up to 24 TB/s and over 2,000 tokens/s when decoding a 4-billion-parameter multimodal model; the company states these are calculated, not measured.
🌍 The architecture targets the 'memory wall' of LLM inference: in the decode phase, each new token requires re-reading weights, and the industry (HBM3E, Groq LPU, Samsung LPDDR5X-PIM) solves the problem by expanding channels. If uHBM survives tape-out, it will become a competitor to HBM in the inference memory market. The first engineering generation is already in the pre-fab pipeline, and the SpinPU-ED01 chip has passed third-party verification: 0.105 TB/(mm²·s) bandwidth density, end-to-end model inference, and 24 hours of stable operation.
👤 For end users of inference services, the platform's success could mean faster and cheaper decoding of large models. This is currently at the architecture level: timelines, process nodes, and prices have not been announced, key figures are calculated, and the first confirmation will be the tape-out results.
Source 1: http://laoyaoba.com/n/1097291 Source 2: https://news.qq.com/rain/a/20260831A07BY700