IT Home, November 12 — Technology media Wccftech posted a blog yesterday (November 11) stating that AMD has introduced the MEXT memory solution for its EPYC server processors, using software-coordinated management of DRAM and NVMe SSDs to reduce memory configuration costs while maintaining application performance close to that of full-DRAM solutions.
IT Home reported in June this year that AMD announced the acquisition of memory optimization company MEXT to address the growing memory bottlenecks in data centers.
MEXT's approach is not to make flash memory as fast as memory itself, but to position data optimally: moving infrequently accessed "cold data" to flash, and using AI to predict access patterns, proactively bringing data likely to be used back into DRAM to mitigate the impact of flash access latency. The solution includes a Linux driver, an AI engine, and a memory monitoring tool called RAM View; according to the developers, the AI engine requires no GPU and uses only a single EPYC processor core.
AMD's tests showed that after running Spark TPC-DS on an Azure three-node cluster with flash handling half the memory capacity, approximately 99% of the full DRAM baseline performance was retained; in the Synopsys PrimeTime chip design software test, approximately 95% of performance was retained even with DRAM halved.
Under similar performance conditions, this solution is expected to reduce DRAM requirements for EPYC servers by 20% to 75%, alleviating pressure from memory supply shortages and expansion costs. AMD has already deployed this technology in its own IT infrastructure.
According to reports, the first 145 servers are expected to reduce DRAM procurement costs by approximately $4.9 million (IT Home note: equivalent to about RMB 32.77 million at current exchange rates), lowering memory expansion costs by 83% compared to an equivalent expansion solution relying entirely on DRAM. The company plans to scale deployment to 2,000 servers by 2027 and collaborate with Hewlett Packard Enterprise (HPE) to promote the system to enterprise customers.
In terms of AI applications, AMD states that with a configuration featuring equal parts DRAM and flash memory, the cost-performance ratio for certain multi-agent and memory-intensive applications can improve by 70% to 90%.