Primemas nears CXL memory mass production with Micron

Primemas Inc. is stepping up its collaboration with Micron Technology as the semiconductor startup moves closer to mass production of its compute express link (CXL), pooled-memory technology for AI data centers.

The company plans to begin mass production of JBOM, short for “just a bunch of memory,” with Micron in the second half of this year, executives said Thursday. The system combines multiple CXL memory cards into a larger shared pool.

The two companies are also working together on Abaco, a large-scale shared-memory project for Pacific Northwest National Laboratory, a US Department of Energy research lab.

CXL is a high-speed connection technology that allows processors and accelerators to access memory beyond what is installed in an individual server. Pooling that capacity can help data centers use memory more flexibly across CPU- and GPU-based systems.

The technology is drawing attention as AI servers require increasingly large amounts of costly memory. Pooled-memory systems allow capacity to be shared across computing systems rather than remaining tied to individual machines.

“Our relationship with Micron is not simply about buying and selling products,” Primemas CEO Park Il said at a media roundtable at the company’s research and development center in Seongnam, Gyeonggi Province.

“We are working very closely together, both strategically and technically.”

Founded in January 2023, Primemas is a fabless semiconductor startup headquartered in Santa Clara, California, with its R&D center in Seongnam.

“With Micron, the partnership has taken shape around JBOM and moved into the mass-production stage,” said Jeon Duk-ho, vice president of technology strategy at Primemas.

The partnership gained further visibility at the Future of Memory and Storage conference in Santa Clara earlier this month, where Primemas was mentioned during a Micron keynote. The companies are also jointly promoting the technology to prospective customers.

Their cooperation extends to Abaco, a rack-scale system designed to support scientific AI, high-performance computing and other data-intensive workloads at the national lab.

Backed by the DOE’s Advanced Scientific Computing Research program, Abaco is designed to provide more than 100 terabytes of CXL-based shared and pooled memory using Micron DDR5. Primemas’ CXL add-in cards are scheduled to be shipped to Micron in September for testing, qualification and benchmarking in the system.

Jeon said the latest FMS marked a more concrete stage for the project, with the Abaco name, its deployment at the national lab and DOE funding publicly disclosed.

Alongside JBOM, Primemas is developing SLiM, or switchless pooled memory, its architecture for pooling memory capacity without requiring a separate CXL switch. The company is targeting applications including AI inference and key-value cache offloading.

Park said Primemas expects to begin generating revenue in the fourth quarter and reach at least $200 million in cumulative sales by the end of 2027. Unlike a conventional fabless chipmaker, the company sells not only controller chips but also cards, boards and system-level products.

Primemas ultimately aims to position itself as an AI data infrastructure provider, with pooled memory emerging as a new layer in computing architecture between DRAM and storage.

“I believe pooled memory will become a new layer in computing architecture,” Park said. “The market could eventually grow to rival the size of today’s DRAM market.”