AMD: Where we're going we don't need (as much) DRAM

AMD is promoting a new technology called Mext, which promises to deliver up to 99% of performance with half the DRAM requirements in their Epyc servers. This is achieved by using AI models to proactively load memory pages into cheaper NVMe storage before the operating system or applications even realize they need them, similar to a smarter swap file.

The technology was acquired from a company called Mext in June and has already saved customers $5 million in DRAM costs on 145 deployed systems. Smerdon, the CVP of business development at AMD, estimates that by replacing costly RDIMMs with flash storage, enterprise server costs could be cut by up to 83%. Mext employs machine learning algorithms, specifically a transformer derivative model, to predict memory access patterns and offload infrequently accessed pages to flash storage.

While flash is cheaper and more power-efficient than DRAM, it is also much slower. To mitigate this, Mext ensures that offloaded pages are back in DRAM before they're needed. EDA and content creation applications are the initial target, but Mext is also being used for latency-sensitive applications like databases and AI pipelines.

In tests, Mext achieved 95% of DRAM performance even with half the memory offloaded to flash, and 99% of performance with 25% DRAM and 75% flash. AMD plans to integrate Mext into its broader enterprise portfolio and will offer it as a license sold through OEM partners.