Kepler Computing Unveils AI Memory Solution to Address Bottlenecks
Chip startup Kepler Computing announced its new memory technology on September 10, aiming to solve AI semiconductor industry challenges in memory power, bandwidth, and cost.
Kepler Computing, a chip startup founded in 2018, officially unveiled its new memory technology on September 10. The company's objective is to address key bottlenecks in the AI semiconductor sector related to memory energy consumption, bandwidth, production capacity, and cost.
The proposed solution offers bandwidth approaching that of SRAM and superior capacity compared to SRAM or HBM, without requiring EUV lithography equipment. The technology utilizes 3D stacking integration, likely based on FeRAM (ferroelectric random-access memory). Kepler Computing expects to have samples ready by the end of 2026 and expand production capacity by 2027.
Looking ahead, Kepler Computing plans to leverage material innovations to overcome limitations in current CMOS technology, aiming to enhance chip energy efficiency and production capabilities through lower operating voltages.
The company is currently producing prototypes using GlobalFoundries' 28nm capacity in Singapore, with plans for future manufacturing in the United States. Kepler Computing stated that the process of modifying existing foundries for this novel memory production took eight months, a third of the typical timeframe.