TL;DR
This paper presents a novel neural network decoder for quantum error correction that can run in real-time on an FPGA, achieving high accuracy with deterministic low latency. This breakthrough addresses a critical bottleneck in building fault-tolerant quantum computers by enabling fast and reliable error detection, and the design is extensible to various quantum code types.
What Problem Does This Paper Address?
Building a fault-tolerant quantum computer requires robust quantum error correction (QEC). A key challenge in QEC is "decoding" – processing syndrome measurements from physical qubits to identify and correct errors. For superconducting qubits, this decoding must happen extremely quickly, within roughly a microsecond per round, and with predictable latency. If the decoder can't keep up, errors accumulate, rendering the computation useless. While existing algorithmic decoders are fast but less accurate, neural network decoders offer higher accuracy but typically run on general-purpose hardware with variable latency, making them unsuitable for real-time QEC. This paper tackles the gap of achieving both high accuracy and deterministic, low-latency decoding on specialized hardware.
Methods
The authors co-design a sparse neural decoder with its hardware implementation on a Field-Programmable Gate Array (FPGA). They adapt a previously developed convolutional neural network architecture, making it "sparse" to fit within the FPGA's computational budget. This involves:
- Hierarchical Sparse Decoder: A convolutional backbone processes spatial information, followed by a lightweight recurrent layer to track logical states. Sparsity is introduced in the channel dimension and through low-precision arithmetic (INT8 and INT4) to reduce computational demands.
- FPGA Implementation: The neural network is mapped onto an AMD Versal HBM FPGA as a fully unrolled, spatial-dataflow pipeline. This means each part of the network is dedicated hardware, processing data in a continuous stream with data-independent control flow, ensuring deterministic latency.
- Extension to qLDPC Codes: The design is extended to high-rate quantum Low-Density Parity-Check (LDPC) codes, which have irregular connectivity. They develop an "affine activation-banking scheme" to efficiently access neighboring checks within a single clock cycle, making these codes also mappable to a single chip.
- Hybrid Arithmetic: They leverage a combination of DSP blocks (for efficient multiply-accumulate operations) and distributed arithmetic (for shift-and-add operations) to optimize the performance of different parts of the neural network.
Datasets
Not discussed in the paper. The evaluation is based on the performance of the decoder on simulated quantum error correction codes (surface codes and quantum LDPC codes).
Key Insights and Ideas
- Neural Accuracy Meets FPGA Speed: The paper demonstrates that high-accuracy neural decoders, previously confined to accelerators with variable latency, can be successfully implemented on FPGAs for real-time, deterministic QEC.
- Hardware-Software Co-design is Crucial: Achieving this feat required a tight integration of neural network architecture design (sparsity, low precision) with hardware implementation strategies (spatial-dataflow, unrolling).
- Feed-Forward Architecture Enables Scaling: The feed-forward nature of the neural decoder is key to scaling to larger quantum codes by simply pipelining multiple FPGAs, without sacrificing throughput or introducing significant latency penalties, unlike iterative decoders.
- Irregular Codes are Mappable: The novel banking scheme shows that even complex, non-planar quantum LDPC codes can be efficiently mapped to hardware, opening doors for more qubit-efficient QEC.
Key Results
- A complete distance-11 surface-code decoder fits on a single AMD Versal HBM FPGA.
- The decoder sustains a microsecond-per-round budget at deterministic latency, crucial for real-time QEC.
- The design is extended to high-rate quantum LDPC codes, specifically targeting the [[144, 12, 12]] Gross code.
- The FPGA implementation utilizes 50.6% of LUTs and 38.0% of DSP58 resources, demonstrating efficient resource usage.
Limitations and Future Work
- Memory Experiments Only: The current results focus on memory experiments; decoding through logical operations (like lattice surgery or transversal gates) is a future direction requiring changes to the convolution graph and routing.
- Outer RNN Prototype: The outer recurrent neural network (RNN) is a floating-point prototype that contributes significantly to the current footprint. Future work will involve mapping this to hardened floating-point DSPs for further optimization.
- Scalability to Larger Codes: While the feed-forward design enables scaling across devices, the paper acknowledges that larger codes will require more FPGAs and inter-device communication, which needs careful management.
Summary updated