The Google Quantum AI team and collaborators has posted a new paper on arXiv showing results of their latest error correction code efforts showing significant results with an improvement in the logical error rate versus the physical error rate by over a factor of two different surface code implementations on their superconducting quantum processors. This shows definite progress and meets the first of the three challenges that GQI has posed for achieving a truly useful fault tolerant computer, what we call a “Sputnik moment”.
Sputnik – On 4 October 1957 the Soviet Union launched Sputnik 1 into orbit. This grabbed headlines around the world, radically raised the profile of this important new technology and kicked off the Space Race.
GQI believes that the three challenges which need to be met for creating a useful quantum computer are the following:
- Challenge 1: Code scaling for systematically suppressing logical errors
- Challenge 2: Universal fault-tolerant circuits with realistic clock times
- Challenge 3: Platform capable of scaling to commercially relevant size
In the paper, they provided data using their latest improved 105 qubit processor to create a distance-7 code using 101 qubits and a distance-5 code with about 50 qubits. With the distance-7 code they were able to achieve a logical error per cycle of 0.143% ± 0.003%, a suppression factor of about 2.14 ± 0.02. The code also increased the qubits lifetime by a factor of 2.4 ± 0.3 over the best physical qubit. What’s more, as shown in the chart above they showed the code does scale with lower error rates as the code scaled from distance-3 to distance-5 to distance-7. In addition, they performed an analysis that showed they could achieve a distance-27 code that achieves a 10−6 error rate with 1457 physical qubits.
Other notable points mentioned was that the performance showed good consistency with operation lasting several hours and up to a million cycles without any degradation in performance. However, the team also noticed large error bursts that occurred approximately once per hour which would represent a noise floor that needs to be understood and mitigated for a full fault tolerant machine.
Perhaps an even more impressive portion of this paper is the description of their work in developing real-time syndrome decoders. These are classical software and hardware that receives the output of the measurement qubits and determines if an error has occurred and, if so, how to correct the error it. The challenge in creating these is to make them highly accurate, low in latency, and high in throughput. This function needs to be performed quickly because the qubits have a limited lifetime and also it can greatly affect the overall runtime of a program that could have million of gate operations. Even though an individual physical quantum gate may be fast, the logical cycle time to detect and correct an error can be orders of magnitude slower. As described in the paper, this team achieved a logical cycle time of 1.1 microseconds while the physical CZ gates have a 42 nanosecond delay on the 105 qubit processor. The paper reported work on two different decoders. The first is a neural network decoder that uses reinforcement learning optimization to fine-tune its operation against processor data. The second is a harmonized ensemble of correlated minimum-weight perfect matching decoders augmented with matching synthesis.
There aren’t very many companies that have described sophisticated error decoder solutions at the level of the ones described in this paper. But we did report last month on Riverlane which is also working in this area with a product technology they call Deltaflow.
But now the Google team has to work on resolving the other challenges that we mention above. Although we have seen several announcements of error correction circuits from several different organizations, they are missing one key thing, support for a universal gate set as described in our Challenge 2. It is not sufficient to achieve quantum advantage if an error correction architecture only implements a set of the Clifford gates (gates including H, S, CNOT, and others that can be generated by combining these together). A universal gate set that can perform any quantum calculation requires an additional gate, a non-Clifford gate such as the T-gate. Future work will need to show how these non-Clifford gates can be implemented using concepts such as T-state distillation factories or some other means.
And Challenge 3 requires scaling up the error correction so that it can provide thousands of logical qubits that will be required to run real-world fault tolerant algorithms. We should note that the distance-7 code that was described in the paper utilized 101 physical qubits out of the 105 present on the chip to create one logical qubit. Sometimes, a technical approach works when you try it on a small scale, but it falls apart when you try to implement it on a larger scale. So researchers need to demonstrate the scalability of an error correction algorithm to make sure it is still practical at larger sizes. Problems that could arise include issues with crosstalk, a slow-down of gate speed, connectivity limitations, cooling, etc.
So we are still waiting for a team to show that can surmount the three challenges mentioned above to achieve what we call the “Sputnik moment”. But in the meantime, the paper demonstrates another welcome step towards the end goal of a fault tolerant computer. Still, many more steps will be needed in order to make fault tolerant quantum computing a reality.
You can access the full pre-print paper posted by the team on arXiv here.
August 28, 2024
