The CMM-Ax, developed by SK Hynix in collaboration with Marvell [SK Hynix]
The CMM-Ax, developed by SK Hynix in collaboration with Marvell [SK Hynix]

SK Hynix has unveiled a CXL (Compute Express Link) memory module capable of handling computation directly, developed in collaboration with global system chipmaker Marvell Technology.

CXL is widely cited as the next-generation AI semiconductor to follow HBM. SK Hynix is moving beyond simply expanding memory capacity by adding computational functions to prepare for the era of AI inference, positioning itself to capture the market ahead of the technology's full-scale emergence.

SK Hynix recently published a CXL technology white paper detailing the results of its collaboration with Marvell — a company with strong capabilities in custom semiconductor design for AI data centers — to introduce the CMM-Ax, a CXL-based processing-near-memory (PNM) memory module.

The two companies are also presenting their CMM-Ax findings together Thursday at FMS (Future of Memory and Storage) 2026, a leading global memory technology conference being held in Santa Clara, California.

CXL is a next-generation interconnect technology that enables CPUs, GPUs and memory chips to communicate seamlessly. While HBM attaches ultra-fast memory directly to a GPU, CXL flexibly connects and expands memory across an entire server system.

The CMM-Ax combines SK Hynix's high-performance DRAM with Marvell's CXL-based intelligent PNM engine, Structera A, which enables more efficient memory use within AI data centers.

The CXL controller integrates 16 high-performance data center CPUs from Arm, the world's largest semiconductor intellectual property company. That enables in-memory computation at bandwidths of up to 200 gigabytes per second.

SK Hynix hopes to supply the CMM-Ax to next-generation AI data centers. The collaboration was driven by the need to resolve the memory bottleneck that has become one of the defining challenges of the AI era.

As AI shifts from training to inference, large language models store previously computed information in a KV cache (Key-Value Cache) to improve efficiency. That eliminates the need to repeat identical computations, but it also causes KV cache capacity to balloon rapidly.

The CMM-Ax addresses this by processing KV cache operations within the memory domain rather than on the GPU, significantly reducing data transfer latency. Performance improvements were confirmed in benchmark tests against a single GPU.

When SK Hynix and Marvell applied the CMM-Ax to a GPU-based backend system (NELSSA), it achieved throughput up to 5.5 times higher than a single GPU and 3.6 times higher than a dual-GPU configuration. The companies said the module can deliver high-performance inference in long-context LLM environments. The validation was conducted with 512 gigabytes of capacity per device.

The semiconductor industry expects CXL to follow a growth trajectory similar to HBM, with some observers calling it the "next HBM" in the AI chip market. Samsung Electronics and SK Hynix, the two dominant players in Korea's memory chip sector, are both deep in technology development aimed at commercialization.

CXL can deliver performance comparable to DRAM while significantly expanding capacity, potentially cutting the cost of building and operating AI data centers. According to market research firm Yole Group, the CXL market is valued at $2.1 billion this year and is projected to grow to approximately $16 billion by 2028.

"Unlocking AI's full potential requires intelligent memory solutions that combine large capacity with high-performance computing," said Joo Young-pyo, executive vice president at SK Hynix. "The successful validation of CMM-Ax is the fruit of a powerful synergy between Marvell's system-on-chip expertise and SK Hynix's leadership in AI memory."


jeongwan@heraldcorp.com