IBM uses an AI agent factory to speed up error correction decoder design 52x

[HardyBoy_WCW]

IBM Research has used AI agents to speed up the hardware design of a quantum error correction decoder, and Quantum Zeitgeist has written up the results from IBM's blog. The approach uses what IBM calls an agent factory, where a coordinating AI agent directs many worker agents running in parallel. It is a neat example of AI helping to build quantum computers

The target was the Relay-BP decoder for IBM's [[144,12,12]] bivariate bicycle code, the qLDPC code at the heart of its fault tolerance plans. The agents explored 100 design hypotheses for each configuration, looking for ways to make the decoder hardware faster

The results are interesting on cost as well as speed. A setup using open weight frontier models, served through IBM's open source llm-d inference framework, achieved a 52 times speedup in high level synthesis estimates for about $26 of inference. It used around 241 million tokens at a blended cost of roughly $0.11 per million, helped by a 98 percent prefix cache hit rate

A setup using only proprietary frontier models reached a bigger 204 times speedup, but cost $299, about 11.5 times more. A hybrid configuration was also tested. IBM is careful to note that each configuration ran only once, so the results are not a firm ranking of the models. IBM Fellow Ruchir Puri is quoted in the post

Decoders need to be extremely fast to keep up with quantum chips, so optimising their hardware matters a lot. Using AI agents to explore designs cheaply and quickly could speed up the engineering side of quantum computing considerably. It is also a nice demonstration of how far open models have come

Is AI assisted chip design going to become standard for quantum hardware? And does the cost difference between open and proprietary models surprise you?

IBM Achieves 52x Design Speedup With AI Agents Tackling Quantum Error Correctionquantumzeitgeist.com

[NeonTundra]

Hitting 98 percent on the prefix cache is a nice technical detail. Reusing context across agents saves a lot of money

That is the kind of optimisation that makes agent factories practical. Agent workloads repeat a lot of context. Caching is where most of the savings come from. It is worth more attention than it gets

[Attention Curtis]

Using AI to design hardware for quantum computers that might one day help AI. There is a nice loop in that

The two fields are becoming deeply connected. Strange times, and exciting ones

[Sandra93]

Fifty two times faster for $26 is a remarkable return. Hardware design exploration used to take engineers weeks. Agents running in parallel can try far more ideas in far less time

Small hardware teams could use this approach too. Twenty six dollars is less than a takeaway

[Rapid Andrew]

llm-d being open source is good to see. Other labs could replicate this setup

Sharing tools like that speeds up the whole field. Universities could try similar experiments without huge budgets. IBM has been consistent about open tooling. Qiskit set that pattern years ago

[DeBruyne93]

Proprietary models reaching 204 times is the bigger number, but $299 is still cheap. Both results are impressive. The real lesson is that AI driven design is now affordable for almost anyone

Different budgets will lead to different choices. For a well funded lab, $299 is nothing. For a student, the open option is the obvious choice