Startup Kepler Computing emerges from stealth mode with an approach to HBM memory that does not require EUV lithography
Startup Kepler Computing has emerged from stealth mode after seven years of development, claiming that its 3D stacking approach to increasing HBM and SRAM memory density does not require EUV lithography. The company has raised 468 million USD from investors and received a pledge of up to 245 million USD from the US Department of Commerce.
Kepler Computing, a startup based in San Jose and founded in 2018, has emerged from stealth mode after more than seven years of development and introduced a new approach to manufacturing memory for AI infrastructure. According to the company, it uses a “3D stacking” architecture combined with a proprietary material to increase the density of high-bandwidth memory (HBM) and fast SRAM cache memory (used in CPU, GPU and XPU), without requiring expensive extreme ultraviolet lithography (EUV). According to the company, the technology is also said to work in existing manufacturing facilities.
Kepler Computing has raised 468 million USD over its lifetime from investors including GlobalFoundries, Intel Capital, AMD Ventures, Baillie Gifford and the Gates Frontier fund of Bill Gates. In July, the US Department of Commerce pledged up to 245 million USD to the company to develop a new class of high-performance AI memory in the USA based on 3D and ferroelectric technologies.
Testing is being carried out in collaboration with manufacturing partner and investor GlobalFoundries, in so-called “mini fabs” in Singapore using a 28-nanometer manufacturing process, as well as at the GlobalFoundries manufacturing facility in Burlington, Vermont. The company claims it was able to convert a fab into a “next-generation” facility in eight months, compared with the usual 24 months, and that its SRAM achieves density comparable to 2-nanometer to 3-nanometer chips without using EUV. According to the company, the ferroelectric approach to SRAM enables data to be read and written at a lower voltage thanks to a composite material that the team developed over 35 iterations.
According to CEO Debo Olaosebikan, the company originally planned to develop SRAM first, followed by DRAM and HBM, but demand for an alternative to HBM after the launch of ChatGPT prompted it to start developing both technologies in parallel. The source article was only partially available; further details can be found in the source article.
Why it matters
The shortage of HBM memory is one of the main bottlenecks in building AI infrastructure because adding new manufacturing capacity is costly and time-consuming. If the approach developed by Kepler Computing were validated in mass production, it could, according to the company, enable capacity increases at existing fabs instead of building new ones for tens of billions of dollars – however, this remains a claim by a startup with no verified commercial deployment.
Relevant practical impact
What this means
For a business
If the approach developed by Kepler Computing is validated in mass production, it could, according to the company, ease the shortage of HBM memory that is critical to AI infrastructure and reduce the costs of building new manufacturing capacity, which range from 20 to 40 billion USD; however, this remains an unconfirmed claim by a startup with no commercial deployment.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.