System-on-chip design startup Baya Systems Inc. said today it has closed on $36 million fuding to support its operational growth and accelerate the development of its software portfolio to meet the needs of the emerging “chiplet” economy.

Today’s Series B round was led by Maverick Silicon and Synopsys Inc., and saw participation from existing investors like Matrix Partners and Intel Capital.

So-called chiplets, or SoCs, have become increasingly popular with the emergence of artificial intelligence, which requires extremely efficient data movement and compute density. SoCs combine various semiconductor components, including central processing units, graphics processing units, neural network accelerators and other kinds of chips to create more efficient computing platforms. However, one of the challenges with putting so many different components onto a single piece of silicon is the need for an efficient design that ensures rapid data movement among the various moving parts.

Chipmakers get around this by integrating network-on-chip technology with their SoCs. NOCs are specialized components that can move data around on the chip while minimizing the resources needed to connect them all together.

They extend across SoCs, but the challenge lies with multi-die implementations — where companies want to connect large clusters of SoCs to run more powerful applications and workloads. Simply put, NoCs don’t work so well when they’re required to communicate with other NoCs.

Baya Systems says it has developed a “revolutionary” chiplet-optimized NoC and physical link or PHY interconnect solution that gets around this challenge. Though it’s possible to connect NoCs on individual chiplets via PHY today, doing so is inefficient, the company says.

That’s why it has come up with a new approach, which involves integrating various NoCs into a “unified fabric” that functions like a single, intelligent, global NoC. According to Baya Systems, its technology helps reduce the latency and increase the bandwidth and throughput of NoC-to-NoC communications, while providing greater flexibility in terms of traffic routing in multi-die designs.

Baya Systems founder and Chief Executive Dr. Sailesh Kumar said designers of complex SoCs and chiplets that combine CPUs, GPUs and other accelerators would typically rely on “brute-force” solutions, and claims that’s unsustainable.

“There are too many risks, such as high re-engineering costs, difficulty in scaling, and potentially hitting the market with sub-par metrics,” he said. “Baya’s performance-focused, software-based approach, coupled with our unique transport and modular fabric IP, is designed from the ground up to produce complex multi-die solutions that are correct by construction with a simplified design process.”

Baya Systems says its technology optimizes and simplifies chiplet communications, freeing systems designers to work on the value-added part of their systems. In other words, chip designers can instead focus on making their chiplets even more powerful.

Maverick Silicon Managing Director Andrew Homan said the challenge of AI is not about providing computing power, but facilitating seamless data movement. “The team at Baya Systems is uniquely positioned to fill this critical gap in the industry with WeaverPro, WeaveIP and its other solutions,” he said.

Baya Systems says early adopters of its software include the AI chip startup Tenstorrent Inc., which recently raised $693 million in late-stage funding to fuel its ambitions of rivaling Nvidia Corp. It’s designing more affordable AI accelerators based on a RISC-V architecture that eliminate the need for expensive components such as high-memory bandwidth.

The startup said it plans to announce more partnerships with chipmakers later this year.

Image: SiliconANGLE/Microsoft Designer