Like a virus, generative AI has invaded nearly every facet of technology, including chip and system design. If you ask Cadence Systems senior VP Paul Cunningham, somewhere between 20 and 30 percent of chips under development today are being designed with the help of AI agents, and that number is growing rapidly.
In many ways, Cadence's latest products are among the clearest examples of how generative AI and HPC will coexist not just in industry but the sciences as well.
Over the past year or so, the EDA giant has rolled out a series of so-called "AI Super Agents," including AuraStack for advanced packaging and PCB layout, ChipStack for SoC design and verification, InnoStack for PHY routing, and ViraStack for analog circuit design.
Much like Anthropic's Claude Code or OpenAI's Codex, these harnesses are designed to work with LLMs to automate design and testing flows by orchestrating Cadence's existing EDA, test, and verification tools. This works for much the same reason that AI code assistants have exploded in popularity among software engineers. Code either compiles or it doesn't. Similarly, Cadence has tests to verify and validate its designs will work as intended.
But just like many code assistants, the effectiveness of these tests depends heavily on how closely their training data resembles the inputs and outputs they're fed. Cadence's early agents were largely powered by off-the-shelf LLMs, which limited exposure to something like RTL.
“It’s not the case that these amazingly powerful out-of-the-box foundation models are just going to do everything. They’re the starting point, and you can still specialize,” Cunningham said.
In a bid for greater specialization, this week Cadence announced a partnership with Nvidia and launched its Maestro AI model family. These models were fine-tuned on Nvidia's Nemotron family of models to work with and interpret the results of Cadence's own tools.
Not surprisingly, better domain knowledge tends to translate into better results. If Cadence is to be believed, early testing of its Maestro models showed an "up to" 30 percent increase in output quality and "up to" 5x greater token efficiency versus general purpose foundation models. In other words, post-trained models on Cadence's tools made them both smarter and cheaper to run.
Along with the models, Cadence's Maestro platform also allows for customization based on customer data, although Cunningham tells us this will be offered more as managed service than something customers figure out on their own.
Much like its existing agent harness, Cadence's Maestro models and agents can access the models via API, run in a secure VPC, or assuming you've got the hardware necessary, on prem.
Cadence isn't the only one exploring the use of AI agents to orchestrate HPC workloads. Nvidia has been pushing HPC into a collision course with AI for years now, which makes sense when you consider its GPUs are good for more than training and running AI, even if that’s what most folks are buying them for.
This concept of AI-driven labs is also a core component of the Department of Energy's Genesis Mission, under which we've seen a flurry of new large Nvidia-based supercomputers.
As you may recall, researchers at Sandia National Labs recently demonstrated how AI agents could be used to develop and test new hypotheses in what they described as a self-driving lab, although those tests didn't have the benefit of modern agent harnesses or even LLMs.
Given the limited funding afforded public researchers, it's easy to see why the idea of an AI lab assistant capable of carrying out tasks long after you've gone home for dinner might be attractive.