Kioxia has found a way to overcome the DRAM shortage – fast SSDs will help, but regular PCs won’t

Kioxia is developing solutions to reduce data center dependence on expensive and scarce DRAM—artificial intelligence-related workloads requiring ever-increasing amounts of memory.

Image source: kioxia.com

The Japanese company’s strategy includes the release of PCIe 6.0 SSDs and the introduction of CXL memory expansion technology, although Kioxia has not directly stated that these products will completely replace DRAM. The manufacturer demonstrated the GP1 series SSD with PCIe 6.0 interface and NVMe protocol, designed to provide ultra-high-intensity I/O operations in AI infrastructure. These drivers support direct connection to the GPU, reducing data transfers on the standard system bus for intensive computing workloads.

The company also launched its CM10 enterprise SSD and NX1 series of data center drives that support liquid cooling. The former uses the 10th generation Kioxia BiCS FLASH technology; the latter has an E1.S form factor and supports PCIe 5.0 interface. These products are targeted at data centers where improved performance and thermal performance are important. Moving some workloads to flash memory will help alleviate memory shortages and rising costs affecting global supply chains.

Kioxia intends to use high-speed NAND memory as a drop-in replacement for DRAM, but not a replacement. One of the most important developments is the previously launched series of memory expansion modules XL1 It has CXL support and is designed with low-latency XL FLASH technology. CXL allows the drive to work more closely with the processor, creating additional memory capacity without having to move every byte of data to DRAM. High-speed NAND combined with the CXL controller reduces read latency by more than ten times compared to traditional NAND solutions. XL1 is positioned as a solution between traditional DRAM and traditional SSD in terms of performance and cost.

Partial replacement of DRAM with CXL modules will help double memory capacity and increase system performance by 30%. This is important because NAND is inherently slower than DRAM, which means completely replacing NAND will result in performance degradation for latency-sensitive applications. Kioxia’s strategy is to add DRAM to AI server configurations, rather than eliminate it entirely.