The Signal
PsiQuantum and AMD report microsecond decoder timing on FPGA
Researchers mapped a quantum error-correction decoder to an FPGA and reported a 10.11-microsecond latency for a distance-8 block. They estimate that 6.59 microseconds is possible at a higher frequency, though the full design has not yet been mapped at that speed.

A transformer-based quantum error-correction decoder has been mapped to a programmable chip at a latency of 10.11 microseconds for a distance-8 logical block. Researchers from PsiQuantum and AMD describe the implementation, called PRISM, in a preprint submitted on November 10.
Source arxiv.org
They estimate that a faster version could reach 6.59 microseconds per block on an AMD VP1902 field-programmable gate array, or FPGA. That estimate rests on individual network layers implemented at the higher frequency; the full network has not yet been mapped at that speed.
Decoding with erasure information
The work targets fusion-based photonic quantum computing, where photon loss and probabilistic fusion measurements can produce errors known as erasures. PRISM takes the known locations of those erasures as input alongside information from error checks, and is designed for protocols subject to both erasure and Pauli errors.
A decoder uses measurement outcomes from an error-correction protocol to determine a logical correction or update the record used to interpret later measurements. The paper represents the decoding problem as a graph connecting checks, error mechanisms and logical observables. Its transformer exchanges information along graph edges using sparse attention.
In the hardware study, the researchers considered periodic logical blocks with distance 8 in both space and time. The authors describe the problem as roughly equivalent to decoding eight syndrome-extraction rounds of a distance-8 surface code. The reported latency runs from the first input of a decoding block to the complete output from the FPGA, including input loading.
Two speeds, different evidence
The team mapped the complete network at 250 MHz. For the reference configuration, the paper's table reports a 10.11-microsecond latency and a throughput interval of 2.39 microseconds per block.
At 400 MHz, the researchers estimate a latency of 6.59 microseconds and an amortized decoding time of 1.56 microseconds per block. The paper says the full mapping at that frequency is future work: the higher-frequency result is based on individual message-passing and perceiver layers rather than the complete network.
Latency and throughput describe different parts of the decoder's operation. The authors say the design overlaps work on successive blocks within the accelerator, allowing a new block to be processed before all work on an earlier one has finished.
Accuracy and hardware trade-offs
The paper also reports logical-error-rate simulations for several fusion-based protocols. The protocols included surface and color codes at distances up to 24, with an error model involving both erasure and Pauli noise. For each fusion network and distance, the team trained a machine-learning decoder.
The researchers report that PRISM matched the chosen benchmark decoder across the tested error domain and outperformed it at lower error rates. The reference method varied by decoding problem: the study used matching for the six-ring network, an integer-programming-based decoder for other networks and a belief-propagation-based method for the bivariate bicycle networks.
The hardware-oriented model uses ternary weights and two-bit activations to reduce the cost of its arithmetic. In simulations of the distance-8 linked loopy diamond block, the quantized models were less accurate than the reference decoders in the full-precision comparison, the authors report, while offering reductions in decoding time and hardware cost.
The team also describes a possible modular approach in which a larger decoding problem is divided into smaller ones and their outputs combined. That approach would require attention to buffering and communication between the decoding problems, which the authors leave for future work.
Sources
Related stories
- Preprint details FPGA platform for testing QEC decoders
- Qblox and AMD report microsecond quantum-feedback tests
- Preprint finds photon mismatch can lower photonic correction thresholds
- Paper Reports Sub-10⁻⁵ Bounds for Optical Bell Resources
- ORCA PT-2 data yields larger malware-model refinement gains in study