A research team has built an AI agent that redesigns a processor core with no human help, and the results are large. The system, called MiracleGro, improved an open-source out-of-order CPU by 85% in instructions per cycle and 27% in clock frequency, which adds up to a 2.34x wall-clock speedup on a standard benchmark. The work is a strong example of an AI chip design agent doing real microarchitecture work.
What MiracleGro does
MiracleGro is an agentic system built on large language models. It starts from the Berkeley Out-of-Order Machine, known as BOOM, a leading open-source out-of-order superscalar core. From there it works through the design loop on its own.
- It reads cycle-level RTL behavior and synthesis results.
- It finds performance bottlenecks.
- It turns architectural ideas into RTL code.
- It weighs cycle-exact performance, chip area and critical path together before accepting a change.
No person steps in during that process. The output is a modified BOOM design that the researchers compare with the human-optimized version.
The numbers
The team tested on SPEC CPU2017 with full reference inputs, a workload suite that is slow and demanding to run. On the full intspeed benchmarks, which cover 30 trillion instructions, the agent’s design showed:
- an 85% gain in instructions per cycle (IPC)
- a 27% higher clock frequency in the SkyWater 130nm process
- a combined 2.34x speedup in wall-clock time
IPC and frequency multiply. A design that does more per cycle and also runs at a faster clock finishes the same work sooner, which is how the two gains compound into the headline figure.
One caveat applies. SkyWater 130nm is an older, open process, far from the leading edge that commercial chips use. The results show the method works in that setting. They do not say how the same gains would carry over to a modern node.
Why small teams are the target
The researchers frame the project around a resource gap. Improving a microprocessor while meeting ASIC limits on area and frequency takes expertise in computer architecture, hardware design, and performance analysis and profiling. Large industry teams can staff performance modeling, RTL design and ASIC implementation in parallel. Small groups of researchers and engineers can’t, and they struggle to keep pace with advances in processor microarchitecture.
An agent that can do the profiling, the thinking and the RTL writing changes that equation. A lab with a few people and access to frontier models could in principle explore design ideas that used to need a much bigger team.
Why BOOM is a useful test
BOOM is already tuned by people, so it isn’t an easy target. Beating a human-optimized baseline by this margin is a harder claim than improving a naive design. The use of full SPEC CPU2017 reference inputs also matters, because short or simplified workloads can hide problems that long runs expose.
The researchers describe the result as evidence of the opportunity in using frontier models to improve RTL-driven platforms. They position it as a tool for implementation-grounded microarchitecture research, where ideas are tested against real hardware descriptions and synthesis results rather than rough simulations.
Open questions
A single result on one core leaves a lot unanswered. The work doesn’t yet show whether the approach holds up on other architectures, how much compute each design iteration costs, or whether the agent’s changes pass the verification steps a commercial chip would need before tape-out. Hardware has a high bar for correctness, since a bug in silicon can’t be patched with an update.
Readers should also keep in mind that the results come from a research system under the authors’ own evaluation. Independent replication on other cores would show how general the method is.
What comes next for agent-built hardware
The natural next steps are testing the agent on other open-source cores and moving to newer fabrication processes. Teams working on open hardware will likely watch whether the BOOM improvements can be reproduced and whether the generated designs can be inspected and trusted by human engineers.
If they can, agents like MiracleGro could become a standard part of the toolchain for academic and startup chip projects. The 2.34x figure is the number that will draw attention, but the more lasting question is how much of the architect’s job an agent can take over and still produce hardware people are willing to build.
Source: arXiv.org