❯ AX100 developer completes tape-out and reports matrix-equation solving below 100 nanoseconds at about 80 milliwatts
Measured siliconScience and Technology Daily reported on November 15 that AX100 developer 安纳智芯 completed tape-out in October. The company describes it as the first fully integrated analog chip for one-step matrix-equation solving. Its team reports measured solving time below 100 nanoseconds at about 80 milliwatts, consistent with simulations. The disclosed stage is tape-out and testing.
Solving through circuitsMatrix equations appear in robot control, communications and some AI training algorithms. Instead of executing a sequence of digital operations, AX100 maps an equation into a circuit and reads a result from its balanced physical state. The approach builds on earlier Peking University research led by Zhong Sun into precise, scalable analog solvers, where controlling error while increasing problem size has been a central challenge.
Why precision matteredPeking University’s earlier research explains that matrix inversion is error-sensitive: fast analog circuits historically struggled with precision and scale together. The team combined low-precision analog inversion with high-precision matrix-vector operations, iterative refinement and block methods, achieving 24-bit fixed-point precision experimentally for a 16×16 matrix. That 2025 work underpins engineering development; the sub-100-nanosecond single solve is a separately disclosed AX100 test metric.
Real-time robotics tasksThe team also tested motion control, planning, reconstruction and simultaneous localization and mapping. It reported 150%–500% speed increases for end-to-end control and planning algorithms and 200%–500% for core reconstruction and SLAM stages. These tasks repeatedly update calculations as the environment changes, providing specific insertion points for a solver. Full shared hardware configurations and comparison conditions were not disclosed, so the gains remain those of the team’s particular experiments.
Task-specific testsThe team evaluated second-order optimization using approximately 2.93-million-parameter autoencoder and 2.72-million-parameter Transformer models, reporting reductions of 92% and 56% in training-update consumption. For a 128×128 communications-channel Gram matrix, reported equivalent throughput improved 20–207 times and efficiency 27.6–61.3 times against selected digital ASICs and GPUs. These comparisons apply to particular tasks, algorithms and model sizes.
A mature-process pathThe report says AX100 uses a mature manufacturing process, with additional tests in robot control, planning and mapping. The company wants to turn analog research into hardware usable across applications. Production scale, pricing and commercial delivery have not been disclosed; practical value depends on integrating the solver into existing computing workflows.
▮ SIGNALDedicated analog circuits could cut time and power in selected solving tasks; a market depends on carrying those gains into complete application workflows.