Samsung will compete for the AI ​​memory market with semi-custom HBM4

Attempts by Samsung Electronics to win over a decent portion of the HBM memory chip market from SK hynix have not yet been successful, but the first company is betting on adapting HBM4 chips to the needs of specific customers, which it will begin implementing next year, according to the South Korean media.

Image Source: Samsung Members

Sources say that the development of semi-custom versions of HBM4 for large Samsung customers has already begun, and the company will begin their mass production before the end of 2025. Samsung’s first clients in this area may be Microsoft and Meta✴ Platforms, which are developing their own computing accelerators and therefore need more efficient memory. The base HBM4 is expected to increase data transfer rates relative to HBM3E by 66% to 2 TB/s, and the maximum capacity will increase by a third to 48 GB per stack.

Samsung can equip its HBM4 chips not only with its own neural coprocessors to speed up operations related to the operation of artificial intelligence systems, but also with additional computing units adapted to the needs of specific customers. This feature of HBM4 generation memory will become the main battlefield for competing manufacturers. Samsung also hopes to start producing the regular HBM4 universal type by the end of next year. Quite unexpectedly, the company’s partner in this area will be one of its competitors—Taiwanese TSMC, which provides services for packaging chips with advanced layout options.