Quantum decoders are a core technology in Riverlane’s Quantum Error Correction Stack, Deltaflow. These decoders can be implemented in software or hardware, where the latter is used by quantum hardware companies requiring mature QEC solutions to correct errors in their qubits in real time.
The team has recently integrated a ‘sliding window’ decoding technique into its proprietary Collision Clustering decoder. This method, essentially, allows arbitrarily long memory experiments to run in a hardware decoder for the first time.
This is an important step forward for real-time decoding – our streaming quantum decoders can process continuous streams of measurement results as they arrive. There is no need to collect all measurement data and wait until after an experiment has finished, and there is no limit on how many rounds of data may be streamed into the decoder.
Furthermore, this recent work also improves the speed of our Collision Clustering decoder and provides better resource scaling.
This work pushes us further along The Riverlane Roadmap, where we are targeting one-million error-free quantum operations by 2026, with intermediate releases every year. The Collision Clustering decoder is a component of the current Deltaflow release, Deltaflow 1.
The Collision Clustering decoder provides fast, resource-efficient and scalable quantum decoding as part of a full-stack QEC solution. You can read more about how we’re developing Deltaflow across qubit modalities here.
It is also available in our software decoder suite QEC Explorer, which provides a range of decoders for offline experiments and simulations to help quantum hardware companies get QEC-ready.
Here are some of the features and benefits of our updated Collision Clustering quantum decoder.
Speeding up
We have improved the speed of Deltaflow 1’s decoder using several optimisations. These include:
- Optimising the memory layout of the system,
- Reducing the number of checks between growing clusters,
- Processing data round-by-round.
As a result, the per-round decoding time is approximately half that of our previous version.
This also means we are below the crucial 1µs per round decoding time up to and including distance 27.
Normalised per-round decoding time, compared for two versions of our decoder across several rotated planar surface code distances. Simulations done using 0.1% circuit level noise for d rounds. Highlighted 1μs as target per-round decoding time.
Resource scaling
Our optimisations have improved the resource scaling for the Collision Clustering decoder. Our utilisation of the FPGA, which corresponds to the cost and power requirements of the decoder, now has better scaling and is smaller than our previous decoder at modest sized codes.
For example, for a distance 23 rotated planar surface code the LUT count has reduced from 21,693 to 4,794, and the flip-flop count has reduced from 15,126 to 3,310 (where a flip-flop is one bit of fast memory held close to the computation).
LUT count, compared for two versions of our decoder across several rotated planar surface code distances.
Long-lived memory
We can now deal with arbitrarily long memory experiments. Before, we were fixed to a certain number of rounds in the memory experiment, meaning we could only maintain the state of a logical qubit for a fixed amount of time.
Now, we can maintain the state of a logical qubit for as long as the underlying physical error-rates allow.
To do this, we use a technique known as sliding window. The decoder deals with a "window" of the data at once, and then slides the window of data being processed through the entire experiment, constantly updating the error state of the logical qubit being decoded.
A space-time illustration of a quantum error correction experiment. The decoder instances start processing batches of syndrome data (called windows – more on this above) while the experiment is still running. Each instance/window must wait for the previous logical operation to finish before it can start its task. The relevant response time is from the end of the experiment until the final decoder output is produced. We know we have avoided the backlog problem if we can execute an unlimited number of QEC rounds, while never exceeding a constant maximum response time.
Decoding in real-time is a huge challenge – and this work is another step towards that goal, which is vital to unlock useful quantum computing. If we don’t decode fast enough, we encounter an exponentially growing backlog of syndrome data.
We’re continuing to tackle this challenge and you can find out more about how we’re developing our decoders and our wider QEC Stack, Deltaflow, here.