Semiconductor Insiders - Podcast EP365: How Agentrys is Revolutionizing Chip Design with Mark Ren

Episode Date: September 11, 2026

Daniel is joined by Mark Ren, founder and CEO of Agentrys. Mark has 26 years of EDA and AI R&D experience spanning IBM Research and NVIDIA Research, driving design automation innovations that power modern chip design. He received the IBM Corporate Award for contributions to the design closure for high-performance microprocessors.… Read More

Hello, my name is Daniel Nenny, founder of SemaiWiki, the Open Forum for Semiconductor Professionals. Welcome to the Semiconductor Insiders podcast series. My guest today is Mark Wren, founder and CEO of a gentrous. Mark has 26 years of EDA and AI R&D experience spanning IBM research and Nvidia research, driving design automation innovations that power modern chip design. He received the IBM Corporate Award for contributions to the design closure for high-performance microprocessors. At Nvidia, he helped establish the company as a world leader for AI for chip

design and GPU accelerated EDA. Mark has spearheaded the modern GPU accelerated EDA movement driving adoption of deep learning frameworks to unlock performance and scalability for EDA workloads. He also led the first industrial LLM for chip design effort, Chip Nemo. Welcome to the podcast, Mark. Hi, Dan. Good to be here. Mark, you have an impressive background. What drew you to semiconductors and what ultimately led you to start a gentrous? That's a great question. I actually developed my interest in computers since early childhood. I learned to program in Basic on an Apple II at a nine years old.

It was a rare opportunity actually for kids in a second-tier city in China in the 80s. So later on my parents bought me a local manufactured computer. It's called a learning PC. And it comes with an architectural design book. And I found myself interested in reading that book and understand how the computer was built, which actually led me into an electrical engineering major in college instead of just a CS major for college.

And this continually interest in hardware also leads me to start my career in IBM doing chip design rather than with software on the internet company. So at the meantime, I was also fascinated about AI doing my college years when I work on computer vision-driven robotics in my undergrad thesis. And my friend and I also did an autonomous driving system design project to participate in a contest using TI's like DSPs at that time. So these are really, you know, where are coming from?

What makes me think about to do your startup in the space is like last year I saw three things that lead to start. This is a startup. The first thing is we saw agent can significantly outperform model itself and reach like near human capabilities. We have a research work called the Vialcoder that can actually solve 94% of the RTL design problems in the VILEval benchmark. Well, the best AI model itself can only achieve at that time like 60%. That's a big jump with agents. It's nearly a production level for that problem. Then the second thing we saw is the wide and fast adoption of coding agents. It only took a couple of months, coding agent cursor to be adopted company-wide inside

Nvidia. So the third thing we saw is the self-improving loop, right? It's actually working. We had a research where we let LM to evolve sat-solver. And we took the 2024 competition winner, side competition winners, solvers, and that the AI to evolve its algorithm and its code. And in one month, it's able to create a server

that beats the competition winners of the side solver competitions in 2025 by a large margin. So it took one month for AI to improve the solvers that much better than all the experts in the world can do in one year. So we saw all this signals that AI can now make significant production impact and change how cheap is going to be designed. And I believe I have the unique background in both EDA and AI to make impact beyond our research. And that prompt me to start a company.

It's interesting. So I saw you at DAC. Your booth was, the agentress booth was just packed with people, you know, so I spoke with you there. But I also saw you at the recent Hotchips conference at Stanford, you guys exhibited. And we'll get into the unusual way you engaged to attendees this year. It's something I'd never seen before. But first, I remember you presented at Hot Chips two years ago while you're at NVIDIA. What was that experience like?

How did that work help shape Adentris? Yeah. Two years ago, I was invited to give tutorials at Hot Chimdivis. chip. We actually did two tutorials. The first one is on AI for chip design in general, and the other is specifically on agentic AI for chip design. Yeah, it's actually two years ago. We are already talking about agentic AI for chip design. And these presentations were an opportunity for us to share what we learned from Apply AI across the real chip design problems at Nvidia. And the central lesson was that the previous machine-duty or AI for Jupy Design work, the scope is quite narrowed and they are relatively difficult setup. But with the generative AI, especially with its coding and reasoning capabilities, it becomes general purpose and applicable to a border range of two-engineering design problems. And we characterize those problems as coding, know-how assistance, analysis, debug, and optimization,

These are like general problems for many design problems, friend and back end, you name it. So I think last, you know, two years ago, it was still early for the audience actually to talk about Agentic AI because the model capabilities was still not production ready, right, for Chbidab. And that is why I did not get to start my own company two years ago, right? Which I, you know, reflect right now, probably should. But I think looking back, right, the medicine is that the speed of AI improvement exceed my wireless expectations. And building production quality AI is still what really matters.

So I think it further my conviction that what agents is doing now is the rising for the future. Interesting. So what was your biggest takeaway from hot chips this year? Yeah, that's a great question. In my perspective, I think open AI. HALAPUILA presentation made a very strong impression on me. The team described completing RTO execution to tape out in roughly nine months and she delivered a 700 watt inference chip optimized for its workload and have a really

good performance comparable to other big vendors. And I think the D plus was not simply like AI writing RTO. is actually the speed of the full loop, right? So they were able to build a loop that start from workload simulation, workload definition, and then cycle accurate simulation, and RTL generation, verification, physical design, everything in the loop, and they can feedback from the later stage or early stage. And that kind of a cross-boundary loop, right, makes the later stage change,

as well as like performance optimization, possible right they claimed to be able to change their design at the you know last moments when they have to freeze the RTO and be able to save area by like 10 percent that's that's huge so I call this kind of workflow AI-based workflow loop engineering right so this is actually widely used today in software right but in a hardware we're getting to it so I think in Trip design the real advantage really come from the engineering this loop, right? And it has an object here. It has measurable numbers. So you need a harness to connect everything so they can run smoothly and then you can monitor the progress. And the shorter, the more reliable this will become, the faster the entire design organization can learn. So our own hot chip demonstration, right, actually apply the same principle, right? To a risk five out of order chip. It's called the Shang San from the open source domain.

And then we create a loop. And this loop will perform workload analysis and design space exploration, RTR generation, as well as RTL optimization, and then design verification, this is physical design, right? The whole loop to GDS. And in eight hours, right, the flow actually executed the entire loop multiple times. And it's able to reduce the power by 28 percent, improve the performance by 40 percent, and reduced area by 10 percent, relatively to the original baseline.

So it's basically a small scale flow, what the open AI did. And our mission is to make this kind of capability, right, to make it not just be an achievement inside a frontier AI lab, but as a commercial system that every chip design company can deploy on its own environment, with his own tools and knowledge and engineering judgment. We also, you know, on the Hachap, we did not just want to show this pre-builder demo, right? we actually invited the attendees to people who come to our booth to bring us a real design challenge. And we use our product. It's called Agentry Studio to build a working multi-agent system for those challenges during the conference.

Yeah, I don't think I've ever seen a company approach an event quite that way. So you took real design problem from visitors at the show and developed solutions right there in your booth. That's quite brave. You clearly have significant confidence in your technology. What kind of challenges were presented to you at the show? Yeah. For example, one attendee asks us to create a multi-agent workflow for 40 gigabytes per second chip-to-chip link using the interlequin protocol, right, across two 30s lanes. And another attendee asked us to build a process in-memory engine integrated with Pytoch. So basically the first one move from like protocol research through RTR and sign off.

The other move from like AI workload and performs modeling down to a. memory system integration. So both requires agents to work across its traditional engineering boundaries. Okay, so how did the agents approach those problems and what results were you able to show? I mean, I saw you doing this at the show. It was quite impressive. Yeah. For the first project, Interlaken Link, our system is able to build a working multi-agent system outspots. It has 44 agents and 26 tools and it produced a complete design from protocol creating the reference model, generate RTIL, verify, and physical design. And it parts with zero errors and all the smoke test.

And it's running open source EDA tools such as the variator uses open and open STA. And for the processing memory engine, we actually built two flows, right? We got two answers from two independent paths. We have a fast analysis path that basically project about the two two-x end-to-end performance for the GPD2 and the Lama product models and 2.5x during decode, but not too much benefit for BERT model, right, which is the correct answer because the Burt is a computer bottom, not a memory bound. Then another part we did is they take a real cycle-acru simulator, was actually built on the fly, and then we run and we measure about the 2.3x performance on the

memory-bound workload. It's two different paths, but pretty much the same answer. So it kind of will verify the analysis are correct for this problem. And I think the system, right, it's not just to produce some design, but also get some answers. I think that's where the attendees really like to see how quickly a broader request, right, become a workflow that they can actually run and check. That's amazing. Great conversation, Mark. So how can listeners learn more about agentress?

Yeah, we have a website, agentries.aI, and I think you can find a lot of information about our product, Agentry Studio. And you can book a demo on the website directly. We're also going to have a webinar with Sammy Wiki, thanks then, on September 15. And where we're going to show the audience how to build multi-agent systems to go from design specifications to GDS in 30 minutes. And we'd like to also hear about the difficult workflows you design. So I'll improve. Yeah, I'm looking forward to it. So thank you, Mark, and I'm looking forward to having you back for updates on your progress.

This is a very exciting time, and I tell you, Gentris is one of the most exciting companies that I've seen today. Oh, thank you so much. It would be my pleasure. That concludes our podcast. Thank you all for listening and have a great day.