Normal Computing 🧠🌡️

In June, we taped out CN101, the world’s first thermodynamic computing chip. We’re now sharing early bring-up results from the first thermodynamic ASIC, showing how a physics-based approach can enable stochastic, stateful, and asynchronous computation directly in silicon. Also on YouTube: youtu.be/iq3QshQevsc #ThermodynamicComputing #CustomSilicon #ASIC #CN101

Several months ago we published a Verilog simulator built by AI agents in 43 days. The project is still running, with more than 10 agents writing simultaneously to a single shared branch. There are no PRs and no separate worktrees. The agents constantly bump into each other, and velocity holds anyway, because the agents resolving the conflicts are more capable than the conflicts themselves. We treat coordination as a temperature parameter. The project runs hot early, while the architecture is in flux and inconsistency is cheap, then anneals as the system matures and bugs take hours to trace. We call this thermodynamic programming, and the same structure appears in neural network training and collaborative writing. Full piece here: normalcomputing.com/blog/thermodyn…

Jin Lim joined Normal Computing this week as Vice President, Global Strategic Growth & APAC. Most recently, Jin served as SVP of SK Group's AI data center initiative, where he led both the technology and business strategy behind next-generation AI infrastructure. Earlier in his career, he built and scaled organizations at SK hynix, Samsung, Oracle, Couchbase, and EMC, spanning technical and commercial leadership roles across the full stack of modern compute. At Normal, Jin will lead and expand our top five global Tier-1 semiconductor accounts while owning the three-year APAC P&L, including revenue, margin, and operating discipline. He will build and scale our commercial engagement team across the region and synchronize it with our forward-deployed engineering model we have put in place. Jin's ability to operate across both the technical and commercial dimensions of our customer relationships will be critical as we scale. "Jin brings a rare combination in deep tech; the experience building transformative R&D solutions as a seasoned semiconductor veteran, and the track record of someone who has scaled commercial organizations at the highest level. I look forward to working with Jin to support our closest partnerships in physics-based computing and beyond." – @FarisSbahi, CEO & Co-Founder We are excited to welcome Jin to the Normal team.

@thomasahle joined Tim Scarfe on @MLStreetTalk to talk thermodynamic computing, bringing AI into chip design, and how you know an AI-designed chip is actually correct. Watch on YouTube: youtube.com/watch?v=5pieVH…

Solving AI's energy demand is a complex challenge that spans the entire stack, from generation and grid connection to power delivery, cooling, and how efficiently the compute itself runs. Normal is focused on the compute layer: how much intelligence each watt buys. At the Waddesdon Transition Forum, our CEO @FarisSbahi joined the panel "Building the Infrastructure Behind the AI Revolution," alongside James Tyler of @Equinix and Jill Macari of @VEIR_Grid, with James Lockyer of @Microsoft Climate Innovation Fund moderating. The Forum is a gathering of senior investors, policymakers, and energy operators, with Secretary @JohnKerry as Co-Executive Chair. Thank you to Nadav Steinmetz, Dor Bershadsky, Climate First and @GalvanizeLLC for including Normal Computing.

Beautiful morning with @MLStreetTalk in Zurich!

Design verification still depends on engineers manually reading hundreds of pages of JEDEC specs and translating them into testable representations. We've released DRAMBench, an open benchmark for autoformalizing DRAM specs and the paper behind it, presented at the @iclr_conf VerifAI Workshop. Our approach introduces an intermediate formal layer of timed Petri nets that capture device states, commands, and timing constraints in a compact, executable model. From that single representation, verification collateral derives automatically. By Jan Ole Ernst, Dmitri Saberi, Derek Christ, Thomas Zimmermann, Rajath Salegame, Suhaas Bhat, Stanislav Levental, Thomas Ahle, and Matthias Jung, in partnership with @FraunhoferIESE. Both DRAMBench and DRAMPyML are open source, Apache 2.0. Read it here: arxiv.org/abs/2605.00058

At @iclr_conf in Rio? Join us for the Normal Silicon Social on Sunday, April 26 from 6:30 to 8:30 PM at the Copacabana Palace. Request an invite: luma.com/3ioi64wv #ICLR2026

Chip design verification consumes up to 70% of the engineering effort on a project, and a significant chunk of that time goes to one task: reading hundreds of pages of natural language specifications and manually translating them into formal, testable representations. We have been working with @FraunhoferIESE on a better approach. Today we are releasing DRAMBench, an open benchmark that measures how well AI systems can formalize JEDEC memory chip specifications into timed Petri net models. Read more: normalcomputing.com/blog/from-spec…

Normal Computing is sponsoring @iclr_conf 2026 in Rio de Janeiro. Our research team will be on-site presenting papers spanning thermodynamic computing, Normal's physics-based computing architecture, and the AI methods underpinning Normal EDA, our platform for silicon engineering. Come talk to the researchers behind the work. We'll share a preview of what we're launching live at the conference. Follow along for real-time updates in the ICLR app. Meet us at the booth: Thursday, April 23 – Saturday, April 25 9:30 AM – 5:30 PM daily

Our homepage refresh is live. Partnered with more than half of the top 10 semiconductor companies by revenue, we're sharing more about how Normal EDA works and where we're headed. normalcomputing.com

"Normal is one of the most difficult places to work, probably in the world, frankly. You're working at the frontier of multiple fields in an effort to push and create a new field forward." In the final chapter of Inside Normal, @FarisSbahi and @zaqqwerty_ai talk about the founding conviction, the recursive loop between Normal EDA and the Carnot hardware program, and what it takes to build at this intersection.

"Why not be part of a company that's actually part of doing the changing? You get to be part of what's happening. You get to actually change the world." In Chapter 03 of Inside Normal, Craig Churchill and Johann George talk about how business and engineering work together at Normal, what the hiring bar looks like, and why the semiconductor industry's design methodology is changing.

What does it look like to build the product that changes how the world's most complex chips get designed? In Chapter 02 of Inside Normal, Hanna Yip, Max Aifer, and Adam DeHovitz talk about building Normal EDA, our purpose-built AI platform for semiconductors: customer deployments, daily product work, and why EDA innovation is required to make entirely new chip architectures possible.

We're hiring across roles careers.normalcomputing.com

What does it look like to build a new class of computing hardware and a purpose-built AI platform for silicon engineering, at the same time? In the first chapter of Inside Normal, Marc Bright, Pete Vigil, and Brandon Birchall talk about the recursive relationship between our EDA platform and our silicon program, and what it means to work on problems that don't exist in textbooks yet.

Took a stab at the Zen of Agentic Software: Explicit is better than virtual The agent should see what you see You should see what the agent sees Build systems the agent can observe, test, and repair Give the agent good tools Prefer working with text The best tool is often a CLI A CLI is just a sub-agent Tool output should be signal, not exhaust Failure should be loud enough to debug Success concrete enough to verify Leave room for surprise, not for confusion Rely on the model’s intelligence, not its knowledge Make knowledge available. Retrieve just in time Build memory so the agent does not relearn the world Memory helps. Memory lies. Design for both Garbage context in, garbage out Intent is the highest-signal context If a model can’t solve problem with perfect context, it can’t solve problem A longer prompt is not a substitute for data Prompting is always the first baseline If one agent is too slow, consider parallelism Verifiers are great, use more of them Do not sneer at glue. Enough glue, arranged well, becomes product

"In the US we're expecting to have a 49 gigawatt shortfall around 2028...2030 on a global scale". @NormalComputing Co-Founder @FarisSbahi on the future energy requirements for compute: "In the next few years we're going to hit this really hard wall when it comes to the energy requirements of compute". "We need to be thinking about next generation architecture from a somewhat radical point of view... exploring new kinds of architectures that would be considered unconventional". "Designing this kind of silicon and optimising with the level of complexity that silicon has... [is] something very difficult to do with legacy software that hasn't really changed in the last 40 years".

.@FarisSbahi and our CBO Craig Churchill joined @etnshow live to talk Normal's growing presence in Europe, our London hardware team, and what AI means for the future of silicon engineering.

At GTC, Nvidia CEO Jensen Huang highlighted a new equation: Revenue = Tokens per Watt × Total Available Gigawatts. Normal Computing CEO @FarisSbahi says this explains the industry’s push for energy-efficient and workload-specific chips: "Not that many years from now, rather than having a very small number of different chips running all of our workloads, we're going to have hundreds, if not thousands of different chips tailored to each workload and application in the data center."

The industry's answer to the AI power crunch is to find more power. Faris's view: finding more energy is not a durable strategy. The goal is to need dramatically less of it. @FarisSbahi and Craig joined @Furrier and @theCUBE at the @NYSE for their "AI Factories: Data Centers of the Future" series. youtube.com/watch?v=pStVzE…