Anatrix (安納智芯), a Chinese semiconductor company, has developed the AX100, an analog computing chip that solves matrix equations through the response of an electrical circuit. According to a Science and Technology Daily report dated November 15, 2026, the company completed tape-out in October. In company tests, a single matrix-equation solve took less than 100 nanoseconds and the chip consumed about 80 milliwatts. This means that analog matrix-equation solving has moved from laboratory demonstration to the integrated-circuit manufacturing stage.
A matrix equation is a calculation that finds several interrelated unknowns at once. It is used in tasks such as updating parameters in AI training, planning robot motion, and detecting signals in wireless communications. A conventional digital computer finds the solution by combining many multiplications and additions. The AX100, by contrast, uses the physical behavior of an electrical circuit to obtain the solution to the equation directly.
The technology builds on a 2025 study of high-precision analog computing published by Sun Zhong and colleagues at Peking University. The study showed how to refine solutions obtained from low-precision analog circuits to higher precision and how to handle larger matrices. The AX100 can be seen as the development of that research into a fully integrated chip. However, it is necessary to distinguish between the time a circuit takes to reach an approximate solution in a single response and the time needed to complete a calculation to the precision an application requires.
Solving equations by using the equilibrium state of an electrical circuit
The way the AX100 works is easiest to understand by considering the matrix equation . Here, A is a matrix of numbers arranged in rows and columns, and x and b are vectors, or single columns of numbers. An ordinary matrix-vector product computes Ax from a known x. Solving a matrix equation, on the other hand, means giving A and b and finding the unknown x.
In the 2025 study, resistive random-access memory (RRAM) was arranged in a grid, with each element's conductance (a measure of how easily current flows, the inverse of resistance) corresponding to an entry of the matrix. When voltages are applied to the array, the current through each element corresponds to a multiplication, and the merging of currents performs the addition. The circuit that stores the matrix data is therefore also used for the computation itself.
Furthermore, when operational amplifiers and other components are connected in a closed loop, the process in which the voltages and currents in the circuit settle toward an equilibrium state can be used to extract an approximate value of x that satisfies the equation. Xiang Rui, Anatrix's CEO, explained to Science and Technology Daily that because the matrix equation is mapped directly onto the circuit and the result is read out just once after the circuit reaches equilibrium, no iterative calculation is needed.
What becomes unnecessary here, however, is the iterative digital computation that would otherwise be required to reach an approximate solution. To obtain even higher computational precision, the 2025 study used "iterative refinement," which combines solving with the analog circuit and a high-precision matrix-vector product.
Specifically, an approximate solution is first obtained from the analog circuit and substituted back into the original equation to calculate the residual. The residual is then fed to the solver circuit as a new input, and the resulting correction is added to the original solution. By repeating this process, the researchers achieved high-precision computation that is difficult with an analog circuit alone.
Another technique that supports the higher precision is "bit slicing." The RRAM used in the study represents eight levels, or 3 bits, per element. Higher-precision numbers are split across several low-precision matrices, and the results of each calculation are shifted and added according to their digit position to reconstruct the full value. A microcontroller handled control tasks such as adjusting values and performing addition and subtraction.
In this way, the 2025 study obtained approximate solutions and corrections from a single response of the analog circuit and combined them to raise computational precision. The figure of under 100 nanoseconds newly disclosed for the AX100, meanwhile, refers to a single matrix-equation solve. When comparing the original 2025 paper with the AX100 announcement, this difference in the scope of processing is important.
From 2025 high-precision research to a fully integrated chip
The results from Sun Zhong and colleagues were published in the journal Nature Electronics on October 13, 2025. The experimental setup at the time used printed circuit boards carrying an RRAM array and operational amplifiers for the low-precision solve, and a 1 Mb integrated chip for the high-precision matrix-vector product. In other words, it was a demonstration system combining multiple circuits with external control.
In the experiments, a 4×4 real matrix reached 24-bit precision in nine iterations. In an evaluation of matrix inversion using a 16×16 real matrix, the relative error was reduced to about 10⁻⁷ after 10 iterations. The researchers also demonstrated a method called BlockAMC, which splits a large matrix into smaller submatrices and combines solves and matrix-vector products for each. Complex numbers can also be handled by expanding the real and imaginary parts into a real matrix, so the approach can be applied to signal detection in wireless communications.
| Disclosure | Hardware and processing unit | Confirmed results |
|---|---|---|
| 2025 peer-reviewed paper | Iterative refinement combining an on-board solver circuit with a matrix-vector product chip | 24-bit fixed-point precision demonstrated on a 16×16 real matrix; relative error of about 10⁻⁷ after 10 iterations |
| AX100 report, November 15, 2026 | Single solve on a fully integrated chip taped out in October | In company tests, solve time under 100 nanoseconds and power consumption of about 80 milliwatts |
Note: The 2025 precision demonstration and the 2026 AX100 single solve differ in equipment configuration and what was measured.
The 2025 study used an 8×8 resistor array as its low-precision solver circuit. Scaling to 32×32 or larger required designing and manufacturing a dedicated chip for verification. In particular, the wiring of the closed-loop circuit, the current when all rows and columns are driven at once, and noise from parasitic components in wiring and elements were cited as obstacles to scaling up. Integrating the solver circuit and the high-precision matrix-vector product circuit onto a single chip was also listed as a future development challenge.
The AX100 develops these research results into a fully integrated chip. According to Science and Technology Daily, manufacturing does not require a cutting-edge semiconductor process and can use mature fabrication technology. The announcement is notable for moving functions that were spread across multiple devices at the research stage into the manufacture of an integrated circuit.
Note also that the large-matrix performance comparisons presented in the 2025 paper included estimates based on measured circuit response times and power consumption. The comparison matched computational precision to FP32 equivalent and normalized GPU throughput per core. For the AX100 as well, actual processing performance will need to be compared under conditions in which matrix size and computational precision are matched.
Performance evaluations in AI training, robot control and wireless communications
Anatrix says it evaluated the AX100 in three areas: AI training updates, robot control and perception, and wireless signal processing. Solving matrix equations is an important step in all three, but the required precision, processing time and share of the overall computation differ.
The company-side evaluation results reported by Science and Technology Daily are as follows.
| Application | Evaluation target and scale | Reported performance |
|---|---|---|
| AI training | Autoencoder with 2.93 million parameters, Transformer with 2.72 million parameters | Training-update burden reduced by 92% and 56%, respectively |
| Robotics and embodied AI | Motion control and planning; 3D reconstruction and SLAM | End-to-end processing of control and planning sped up by 150–500%; core processing of reconstruction and SLAM sped up by 200–500% |
| Wireless communications | 128×128 channel Gram matrix | Equivalent throughput 20–207 times and equivalent energy efficiency 27.6–61.3 times those of digital ASICs and GPUs |
Note: All figures are company-provided evaluation results, and the measurement targets and comparison metrics differ by application.
The AI training evaluation used a second-order optimization method, which calculates parameter updates using information about the curvature of the loss function. This calculation requires solving matrix equations, so the AX100's analog circuit can be used. The reductions of 92% and 56% presented by the company relate to the training-update processing of each model and do not represent the processing time of training as a whole.
In robotics, the evaluation covered control and planning, which decide motion, and perception, which grasps the surrounding environment. SLAM is a technology by which a robot or other device simultaneously estimates its own position and builds a map of its surroundings. In this evaluation, the core computation was measured for 3D reconstruction and SLAM, while the end-to-end algorithm processing was evaluated for motion control and planning. This difference matters when considering how much a speedup of the circuit alone carries over into the processing time of real applications.
For wireless communications, an evaluation targeting a 128×128 channel Gram matrix reportedly showed equivalent throughput 20 to 207 times, and equivalent energy efficiency 27.6 to 61.3 times, those of existing digital ASICs and GPUs. DIGIT.CN explains equivalent throughput as the amount of processing per unit time obtained by converting the analog circuit's processing into a number of digital operations. That article is based on information from Science and Technology Daily and is not an independent performance verification.
The wireless application was also demonstrated in the 2025 paper. The research team used 256-QAM modulation in a 128×8 MIMO (multiple-input multiple-output) communication system and achieved a bit error rate equivalent to FP32 digital computation after three iterations. This experimental result shows that the approach not only solves matrix equations quickly but also secures the signal-detection precision that communications require.
On the other hand, for the AX100's test result of a solve time under 100 nanoseconds, the size of the matrix tested, the required precision and the measurement scope, including input and output, have not been disclosed. For the reduced burden in AI training, absolute processing times, energy consumption and detailed comparison conditions have likewise not been made clear. The start of mass production and the terms of sale or availability have not been announced either.
To judge the AX100's practical performance, it will be important to state the target matrix size and required precision, and to measure the processing time and energy consumption from input to output of a solution at the required precision. In particular, if the performance of the entire process, including iterative refinement, data conversion and peripheral control, is made clear, it will be possible to assess how far the analog circuit's fast response contributes to efficiency gains in AI training, robot control and communications processing.
