Historically, the main way the semiconductor industry increased chip speeds year over year is by shrinking transistors so more of them can fit on a processor. Cerebras Systems Inc. is taking a different approach: Instead of making the transistors smaller, the California chip startup has made the processor bigger — much bigger.

Cerebras today revealed the Wafer Scale Engine, an artificial intelligence chip 57 times the size of Nvidia Corp.’s flagship V100 data center graphics card. The startup hails it the largest processor ever made.

The Wafer Scale Engine, or WSE for short, is an 8.5-inch by 8.5-inch amber tile (pictured) containing no fewer than 1.2 trillion transistors. That’s compared with 21.1 billion in the V100. The WSE’s circuits are organized into 400,000 processing cores specifically optimized for AI and 18 gigabytes of high-speed onboard memory.

Cerebras is bringing the chip to market as part of a data center appliance that has its own water cooling system to absorb the heat from all those cores. Andrew Feldman, the startup’s chief executive, told Fortune that the machine will provide 150 times the computing power of a server with multiple Nvidia graphics cards. Pound for pound, Cerebras claims its appliance uses 2% to 3% of the space and electricity that would be needed for an Nvidia-based server farm with the same processing capacity.

That efficiency is the direct result of the WSE’s supersized form factor. In a traditional AI environment made up of multiple individual GPUs, the data being processed has to constantly travel between the different chips over relatively slow network links, which creates a bottleneck. That’s not an issue for the WSE because all the calculations are done on the same circuit board.

The idea of centralizing processing operations on one big chip to improve efficiency has been around for decades. However, no company before Cerebras managed to pull it off because of the daunting technical challenges involved in implementing the concept.

One of the biggest obstacles lies in the manufacturing process. Even at the world’s most modern chip fabrication facilities, there’s no way to produce a processor the size of the WSE without at least some of its 1.2 trillion transistors coming out wrong. Defective transistors can normally render an entire processor unusable, but Cerebras has gotten around the issue by building a network into the WSE that can route data around malfunctioning parts.

The chip’s design “takes the fundamental properties of cores, memory, and interconnect to their logical extremes,” Andy Hock, Cerebras’ director of product management, wrote in a blog post. “By connecting everything on-die, communication is many thousands of times faster than what is possible with off-chip technologies like InfiniBand.”

The startup’s technology is attracting serious industry attention. Cerebras has raised over $200 million in funding from investors including early Twitter Inc. backer Benchmark, former Advanced Micro Devices Inc. chief technology officer Fred Weber and Ilya Sutskever, the chief scientist of OpenAI.

Cerebras plans start shipping its WSE-based data center appliance in September. The startup has provided prototype versions of the machine to several early adopters. 

Photo: Cerebras Systems