Skip to content
Quantum.world

The Signal

Preprint details FPGA platform for testing QEC decoders

A preprint describes CHEQ, an FPGA platform that generates circuit-level noise samples and tests multiple quantum-error-correction decoders on-device. In tests on an AMD VCU118, the researchers report 49.2 million shots per second with 20 Helios decoders.

Preprint details FPGA platform for testing QEC decoders

A preprint posted on Nov. 17 describes CHEQ, an FPGA platform that generates circuit-level noise samples and tests multiple quantum-error-correction decoders on the same device. Sechan Park, Kyeongwon Lee, Mundo Jung, Chaebin Jung and Woojoo Lee present it as a way to evaluate decoder logical-error performance and response times with the hardware running as the device under test.

Source arxiv.org

Sampling and decoding on one FPGA

CHEQ combines Sampling Units, a decoder interface, a scoring unit and a RISC-V processor on one FPGA; it generates and scores shots on chip, with only aggregate counters and failure logs crossing the device boundary after configuration. Each Sampling Unit takes a detector error model exported by Stim and uses geometric skipping to sample where error mechanisms fire across a batch of shots.

Rather than test every mechanism for every shot, the unit samples firing positions and amortizes mechanism traversal across the batch. Under independent random draws and exact geometric sampling, the paper says batching leaves the shot distribution invariant while batch size sets a memory-throughput tradeoff.

Tests with Helios

The evaluation engine was implemented in Verilog and integrated with a Rocket RISC-V core, a DDR4 controller and peripherals for an AMD VCU118 FPGA board. The test setup used a 100 MHz clock and Vivado 2024.2.

The workloads were rotated surface-code memory-Z circuits generated by Stim v1.16.0, with code distances 7, 9, 11 and 13 and syndrome-round counts equal to the code distances. The four generator noise parameters were set to physical error rates of 10^-3 and 5×10^-4, and the researchers applied no additional idle noise. The decoder under test was the Helios circuit-level union-find decoder, with its decoding RTL unchanged.

Throughput and timing results

Across 48 evaluated conditions, the authors report a pooled one-sided 95% upper bound of 0.16% on the weighted root-mean-square relative difference in per-mechanism firing rates compared with Stim. At a batch size of 4,096, one Sampling Unit used 4.8%–11.2% of a Helios decoder’s LUTs across the tested code distances.

On the VCU118 at code distance 7 and a physical error rate of 10^-3, two Sampling Units supplied 20 Helios decoders at 49.2 million shots per second. The same configuration completed 10^10 shots in 3.4 minutes without host-side shot generation. At the same physical error rate, the paper reports that a 10^10-shot campaign took 56 minutes at code distance 13.

The paper also compares the Sampling Unit with RISC-V software and a 256-comparator array on the FPGA. For code distances 7 through 13, the Sampling Unit required 3.79–26.27 cycles per shot; the comparator array required 6.1–6.3 times more cycles and 6.1–13.3 times more LUTs.

The evaluation also measured decoder deadline violations: at a physical error rate of 10^-3, the one-round violation rate rose from 4.39×10^-7 at distance 7 to 3.02×10^-5 at distance 13. At a physical error rate of 5×10^-4, that rate rose from 1.43×10^-8 at distance 7 to 1.02×10^-6 at distance 13.

Sources

  1. arXiv
  2. arXiv (full text)

Related stories

Explore

More articles