An ambitious startup has spent over seven years quietly reimagining computer memory architecture. Now emerging from stealth mode, it believes that its novel approach could help mitigate the global memory chip shortage, provided it can successfully scale production of its technology.
Kepler Computing, a California-based startup founded in 2018 by a team of physicists and computer scientists, claims to have developed a new architecture for high-bandwidth memory (HBM). This innovation purportedly addresses some of the bottlenecks in chip supply that have been constraining the computing market.
While chip manufacturers typically rely on the expensive extreme ultraviolet (EUV) lithography to reduce transistor sizes on chips, enabling greater technology integration within the same space, Kepler asserts that its “3D stacking” approach combined with a proprietary material allows for increased density without any dependence on EUV. This method is compatible with existing semiconductor manufacturing plants.
Kepler also claims to have achieved similar advancements in high-speed cache memory, commonly found in CPUs, GPUs, and XPUs. This type of memory, known as SRAM, is located within the chip’s core to minimise data transfer times. In contrast, HBM utilises stacks of DRAM, which act as standalone memory components.
Over the past few years, the company has raised $468 million in funding from prominent investors including GlobalFoundries, Intel Capital, AMD Ventures, the British investment fund Baillie Gifford, and Bill Gates through his private fund Gates Frontier. In July, the US Department of Commerce committed to providing Kepler with up to $245 million to “develop a new generation of high-performance AI memory technology in the US, based on innovative 3D and ferroelectric technologies.”
Strategic Manufacturing Initiatives
Currently, much of Kepler’s testing is conducted in Singapore, where it operates a plant owned by GlobalFoundries, its manufacturing partner and investor, which has contributed $50 million. Over the past two years, Kepler has been developing what it refers to as “mini factories,” where it produces its memory chips in collaboration with GlobalFoundries’ 28-nanometre chips. Additional testing is also taking place at GlobalFoundries’ facility in Burlington, Vermont.
Kepler and its investors assert that the startup’s innovative approach to memory chip manufacturing arrives at a pivotal moment. The global memory chip shortage has reaffirmed two market realities: building new factories is incredibly costly and labour-intensive, and demand for certain types of memory tends to be cyclical. In a world dominated by data centres and artificial intelligence, HBM has emerged as the preferred choice.
Major memory manufacturers such as SK Hynix and Micron have rushed to construct multi-million dollar factories to meet this demand, betting that even when these facilities come online, the industry will continue to seek high-bandwidth memory solutions.
“For a long time, we thought we would work on SRAM first, then DRAM and HBM,” noted Debo Olaosebikan, co-founder and CEO of Kepler Computing. “However, with the launch of ChatGPT in 2022 and the surging demand for an alternative to HBM, we began to develop our roadmap for HBM. Now we are manufacturing SRAM and HBM simultaneously.”
“We didn’t start out thinking this would be a replacement for DRAM or SRAM,” clarified Srini Ananth, managing director of Intel Capital. “We figured the market would dictate that, and now we see there is demand for both.”
Avoiding Traditional Bottlenecks
By sidestepping EUV lithography, Kepler positions itself as one of the few tech startups working to circumvent one of the primary bottlenecks in semiconductor manufacturing. The company claims it can produce SRAM with the same density as 2 or 3-nanometre chips without needing to invest in EUV lithography.
“Kepler’s approach aligns perfectly with our strategy,” stated Ed Kaste, senior vice president of the CMOS business at GlobalFoundries. “It represents a new materials system with multigenerational scalability potential, without the need to build entirely new systems in the factory or invest in expensive lithography equipment.”
Kepler’s primary proposition is that the accelerated computing market should not have to wait for new memory factories to meet demand. Instead, innovative manufacturing methods within existing plants can increase supply.
The startup’s dual approach focuses on enhancing HBM, claiming to have developed a novel 3D manufacturing technique that allows for the integration of more memory chips in a compact space. This proximity between processing cores and memory is designed to reduce energy consumption during data transfers. Ultimately, Kepler aims to manage data transfer in HBM with energy consumption levels comparable to SRAM while maintaining substantial capacity.
Additionally, Kepler has reportedly improved SRAM density using ferroelectric materials capable of reading and writing data at lower voltages than conventional mechanisms employed for processing and storing data in semiconductors. This innovation stems from the development of a new low-voltage composite material that supports the ferroelectric approach.
Challenges Ahead for Mass Production
Sasi Manipatruni, co-founder and chief technology officer of Kepler, explained that the team underwent 35 iterations of compounds before identifying a material class that would facilitate and lower the cost of memory chip manufacturing. “Once we figured out how to solve the physical problem (i.e., the physical limitations of HBM), we came up with a materials innovation that helps to increase the amount of memory that can be situated between chips,” he explained.
In initial implementations with GlobalFoundries, Kepler claims it was able to transform a factory into a “next-generation” facility in just eight months, compared to the standard 24-month timeframe. “Our goal is to leverage existing factories and architectures, pushing them to the limits of physics,” Olaosebikan remarked.
Essentially, Kepler is betting that any additional costs incurred from new materials or the repurposing of existing factories will be significantly outweighed by the estimated $20 billion to $40 billion required to build and equip new facilities with hundreds of millions of dollars worth of machinery.
Kepler Computing still has a considerable journey ahead before it achieves large-scale production, assuming it succeeds in this endeavour. To date, the company has tested its technology on around 2,000 wafers. It plans to ship its first HBM chip samples by the end of this year, ramp up production from Singapore next year, and commence chip production in the US by 2028.
Anticipating Future Developments
Kaste, the GlobalFoundries executive, is optimistic that “the fundamental advancements have already occurred. What remains is to achieve good results on thousands of wafers and millions of devices.”
He pointed out that one challenge posed by a materials system like Kepler’s is its inclusion of iron in its compound. “Iron is a difficult contaminant to introduce into a production plant. Therefore, Kepler’s solution must work on specific equipment or be entirely encapsulated to prevent any escape,” he explained.
“The art lies in keeping that material well contained throughout our production flow,” Kaste argued.
Olaosebikan refrained from confirming the specific elements of the company’s composite material. “What I can say is that we are using a limited number of materials, and some of them are not those typically found in conventional ferroelectrics,” he disclosed.
Kepler Computing is not the only startup aiming to revolutionise the semiconductor industry. Last year, a well-funded startup named Substrate also attracted attention for its novel lithography approach that employs nanoparticles to etch details onto advanced chips. Some industry analysts expressed skepticism regarding the startup’s ambitions, noting the considerable challenge of producing a large volume of chips that meet incredibly stringent specifications, on time and within budget.
The scale of production remains a common challenge when trying to innovate in semiconductor manufacturing, whether through new processes, materials, or a combination of both. “The question is how to overcome the limitations related to contamination, differing materials, and various tools in such a way that the resulting innovation can be scaled and justify the cost,” remarked Austin Lyons, a chip analyst at Creative Strategies, who was not briefed on Kepler’s innovations.
Demonstrating the viability of a new method for memory is one challenge; meeting the historical demand for it presents an entirely different set of hurdles.