Cerebras Systems has secured another U.S. government win for its wafer scale engine chip – which is considered the largest chip in the world.

The company’s chip technology will be part of a research project sponsored by the National Nuclear Security Administration to find future computing technologies to assess the nuclear weapons stockpile.

The NNSA and its partners will assess the chip’s capabilities as part of an initiative to research post-exascale technologies.

The NNSA’s Advanced Simulation and Computing (ASC) program is researching technologies that could be 40 times faster than the upcoming exascale system called El Capitan, which will be hosted at the Lawrence Livermore National Laboratory in Livermore, California.

The U.S. Department of Energy has said that the El Capital supercomputer, which will be the first multi-exaflops system in the U.S., will go online in 2023 or 2024, according to a 25-year ASC accomplishments report published this month. The system is being built by HPE in collaboration with AMD, and Livermore Lab has already taken delivery of three testbed systems.

The Cerebras system could become a part of the Department of Energy’s program to evaluate experimental systems, which is called “Advanced Architecture Prototype Systems.” The first AAPS system, called Astra, was a petaflops-system based on Arm processors. A followup system, codenamed Vanguard-2, is currently being built. Development of the next AAPS system, Vanguard-3, will start in 2026, according to the ASC roadmap.

The post-exascale program is also considering “quantum computing and other novel hardware, computer architecture, and software; the likely trajectory of relevant hardware and software technologies; and the ability of the U.S. industrial base to meet NNSA’s needs,” according to the project page.

NNSA’s goal is to improve the computing capability of the Stockpile Stewardship Program, which involves the assessment of the current nuclear arsenal. The annual program helps the government simulate and evaluate nuclear weapons without real-world testing. Better computing capabilities will improve the testing capabilities.

The Cerebras wafer-scale engine, which has 850,000 cores and 2.6 trillion transistors, was already being evaluated by the U.S. government. The Argonne National Laboratory, which focuses on scientific research, was assessing Cerebras’ chip as an alternative to GPUs for artificial intelligence applications. The NNSA and affiliate labs that include Los Alamos, Sandia and Lawrence Livermore are focused on nuclear stockpiles and other weapons technologies. At Livermore, Cerebras’s first-generation CS-1 machine was integrated into the NNSA’s unclassified Lassen supercomputer in 2020.

“As a startup, it’s extremely difficult to win the trust and get access to the workloads, even to see if you’re good at them,” Andrew Feldman, CEO of Cerebras, told HPCwire.

The NNSA project allows Cerebras to collaborate with a whole new class of customer, Feldman said.

Cerebras’ chips have also been used by private sector companies that include GlaxoSmithKline and TotalEnergies. The WSE-2 chip is finding more acceptance in the private and public domain, and the product has matured as more applications are found and AI models mature.

“In the software world, they say the first thing to do is get in customers’ hands. In the hardware world, it takes us years to build chips and get systems. Then we have to get them in customers hands and learn as quickly as we can,” Feldman said.

Cerebras’ chip has made its mark in scientific and research applications, but it’s not yet in front of corporate customers for everyday use. Google Cloud and AWS have put up instances of their homegrown AI chips for such applications, but Cerebras’ chip is significantly faster, and can be integrated into high-performance computing workflows.

“We’re not going to be an exclusive cloud seller. We’re going to continue to deliver to customers, who, for whatever reason, also want on premise solutions,” Feldman said.

Cerebras would love to partner with major cloud providers, Feldman said, adding “you have to earn that. That takes time.”