Synopsys and Ansys power the future of innovation—connecting silicon to systems.
EDA
System
IP
SolvNetPlus
Training & Education
Gone are the days when chip companies could focus solely on building and selling chips.
Today, they must address a much broader set of demands, and the challenge is especially acute for the growing number of startups seeking a foothold in the multitrillion-dollar AI infrastructure market.
While specialized silicon is often the core innovation on which these startups are built, customers increasingly expect solutions that can be deployed and operated at scale — complete systems that include hardware, software, and infrastructure.
“To build a winning AI chip startup from scratch, it’s not just about the chip,” said Andy Hock, chief strategy officer at Cerebras. “You’ve got to think about the whole solution.”
Navigate AI Chip Development
Your essential guide to overcoming AI chip complexity and achieving successful silicon outcomes from design to deployment.
Taking different paths to market
Hock was speaking at the AI Infrastructure Summit in Santa Clara, California, alongside leaders from two other chip startups, Etched and AheadComputing, as well as Meta, a hyperscale consumer of AI chips. The panelists agreed that turning silicon innovation into a deployable product requires focus, early customer validation, and engineering that extends well beyond the chip.
Each company, however, is forging a different path to market.
Portland, Oregon-based AheadComputing, for example, is developing high-performance RISC-V CPU cores for a growing range of computing systems. According to co-founder and chief architect Jonathan Pearce, the company’s reusable chiplet strategy aims to shorten “the time from innovation to silicon,” so the same building blocks can serve different applications and markets.
“The AI era has kind of rewritten the systems-level view of what a computer can be,” Pearce said.
Etched is taking a more targeted approach, designing its Sohu chip specifically for transformer inference.
“As a startup, you have to be super-focused on the design space you solve for,” said Robert Wachen, co-founder and president of the San Jose, California-based company. “Being very focused in the design space has enabled us to execute faster, and we think the demand for inference has gotten to a point where that makes sense.”
Building the product around the chip
To accelerate time to market and deliver at scale, Etched made two strategic decisions. First, instead of working through a series of smaller test chips, it collaborated with Synopsys to emulate its full-reticle design at full scale. The team ran more than half a dozen complete models end to end on the emulated chip.
“That gave us the confidence to go tape out and go for the full thing,” Wachen said.
Second, Etched took on complete ownership of the surrounding hardware and software stack, including the inference engine, host, and embedded device software. The goal: reducing integration hurdles for customers.
“We co-designed the full system in parallel,” Wachen said.
Before tape-out, Etched shipped boards, racks, and cooling equipment — everything except the AI chips — to customer data centers to get feedback on the system design. The company treated tape-out as a milestone, Wachen noted, not the finish line. The team validated software and used thermal test chips to check the cooling hardware before silicon returned from the fab.
Etched calls this approach “prefetching,” borrowing the computing term for retrieving data before it’s needed.
“You have to parallelize the whole schedule,” Wachen explained. “We come up with ways to de-risk every component as much as possible. Silicon coming back from the fab should kind of be the last thing in the system, and that should enable you to ship as quickly as possible.”
Cerebras took a similar full-system approach with its wafer-scale processor. With more than 4 trillion transistors, the chip is the size of a dinner plate. Deploying it required Cerebras to integrate cooling, power delivery, data movement, and software into a complete system.
“We had to literally invent machines, materials, and new processes to handle and deploy our systems at scale,” Hock said.
Usman Ali, a member of Meta’s infrastructure silicon architecture team, spoke from the customer’s side. He said a novel chip on its own won’t win over a hyperscaler.
“Hyperscalers serve billions of users,” he said. “Unless you have a fully mature system, a hyperscaler will have concerns.”
Left to right: Sandeep Mehndiratta (moderator, Synopsys), Robert Wachen (Etched), Usman Ali (Meta), Andy Hock (Cerebras), Jonathan Pearce (AheadComputing)
Using AI to compress development
Building the chip, system, and software in parallel puts heavy demands on small engineering teams. To keep up, chip developers are applying AI to the engineering work itself. Ali said some chip design tasks that used to take two weeks at Meta can now be completed in a day with AI. Verifying AI-generated work is still a significant challenge, he added, though he expects verification to become more automated over the next two to three years.
Etched uses AI across chip and system development, from kernel design and PPA optimization to engineering telemetry, dashboards, and reporting.
“You start with small, verifiable tasks and increase the scope over time,” Wachen said. “Your agent should be doing something while you sleep.”
For a startup, faster iterations can create a clear advantage.
“If I can accelerate iterative experimentation and close that design loop more quickly, that’s what allows me to burn down risk,” Hock said. “That’s what allows me to get to market. That’s what allows me to win first.”
From innovation to deployment
Growing demand for AI compute is fueling interest in specialized silicon and new system architectures.
“Teams like ours and those at Meta have shown that there is value in chips that are built for purpose,” Hock said. “With this capital inflow and expanding value demonstrated, I think there’s an opportunity to see more and different heterogeneous architectures come into the world.”
The startups that capture this opportunity will need to turn specialized silicon into complete systems that customers can evaluate, deploy, and run at scale. That means developing the chip, software, and system in parallel while managing risk and making the most of limited engineering resources.
AI can help by accelerating iteration and automating more engineering work. Access to proven design technology and expertise can also help startups tackle increasingly complex silicon and system challenges.
Together, these capabilities can shorten the path from a promising architecture to a complete, market-ready system.
Continue Reading
