{"slug": "amd-where-we-re-going-we-don-t-need-as-much-dram", "title": "AMD: Where we're going we don't need (as much) DRAM", "summary": "AMD says its Mext memory-tiering software can deliver up to 99 percent of server performance with half the DRAM by using transformer-derived machine learning models to predict and pre-load memory pages from NVMe flash before the OS or applications need them. Gary Smerdon, former Mext CEO and now AMD CVP of business development, said the technology could cut enterprise server costs by as much as 83 percent, and that across the initial 145 systems it has been deployed on, customers have already saved $5 million in avoided DRAM costs. AMD acquired the technology from Mext in June 2026, and it requires a Mext license alongside an Epyc server.", "body_md": "# AMD: Where we're going we don't need (as much) DRAM\n\n[The Register](https://www.theregister.com)\n\nThe House of Zen's Mext team promises smarter SWAP through the power of, you'll never guess: AI\n\nAMD's latest sales pitch: Up to 99 percent of the performance with half the DRAM, provided you're willing to buy a Mext license alongside your shiny new Epyc server.\n\nThis is made possible by using AI models to load memory pages in and out of cheaper NVMe storage before your operating system or apps ever know they need it — kind of like SWAP or a Windows page file, but smarter.\n\nAMD [acquired](https://www.theregister.com/systems/2026/06/16/amds-mext-buy-shows-how-ai-could-solve-the-ram-shortage-it-created/5257352) the tech from Mext back in June, and according to former CEO Gary Smerdon — now CVP of business development at AMD — the tech could cut enterprise server costs by as much as 83 percent by replacing costly RDIMMs with flash storage. Of the initial 145 systems the software has been deployed on so far, he estimates customers have already saved $5 million in avoided DRAM costs.\n\n\"Conservatively, 60, 70, even 80 or 90 percent for very large memory systems, of the cost of a server can be in memory,\" he explained.\n\n### Everyone's favorite buzzword to the rescue\n\nEven in systems with high memory utilization, where (say) 95 percent of the memory is \"in use,\" a lot of it isn't actively being used, he explained. He cited a Meta research paper that showed that between 50 and 80 percent of memory pages hadn't been touched in five minutes, ten minutes, or even longer. \"Five minutes to Epyc is an eternity.\"\n\nMext technology aims to offload those infrequently accessed pages to flash. The challenge, Smerdon explains, is that while flash may be 50x cheaper per bit, 30x lower power per bit, it's also 500x slower.\n\nThis is why most folks, when provisioning VPCs or servers, generally avoid swapping out to disk wherever possible: it kills application performance and sends service level agreements out the window.\n\nSo for Mext's tech to work, it needed to ensure that any pages offloaded to storage were back in DRAM before they're needed. Kind of like running an errand in the middle of the workday – so long as you're back before your boss needs you again, they'll never know you left your desk.\n\nThe problem was that the \"R\" in DRAM stands for random. Predicting which pages the system was going to need at which time isn't exactly trivial. The solution? (Brace for it.)\n\n\"We realized we could use AI to monitor the system and proactively push those pages back into DRAM, so that as far as the application or the operating system knew, they were there the whole time,\" Smerdon said. In other words: \"We transparently make flash appear as DRAM.\"\n\nSo, \"instead of buying two terabytes of DRAM for your Epyc-based server, you could buy one terabyte of DRAM, and then the next terabyte comes with Mext.\"\n\nTo do this, AMD’s Mext employs a number of [machine learning](https://www.machinebrief.com/glossary/machine-learning) algorithms.\n\n\"The workhorse model is a [transformer](https://www.machinebrief.com/glossary/transformer) derivative model, just like [ChatGPT](https://www.machinebrief.com/compare/chatgpt-vs-claude) or Claude or the other famous models, but instead of predicting, for example, in the case of text, what word is next, what we're doing is predicting what page of memory is going to be required next,\" Smerdon explained. So long as the pages are in DRAM before they're needed, the pages don't even need to be in the right order, he added.\n\n[Training](https://www.machinebrief.com/glossary/training) the models to predict memory access patterns accurately presents its own challenges, which, Smerdon says, makes [pre-training](https://www.machinebrief.com/glossary/pre-training) the models impractical.\n\nInstead, they take a single Epyc core and train the model \"in under one second,\" he explained. From there, the model is further refined and retrained over time until it is able to predict accurately where the pages need to be and when they need to be there.\n\nAll of this training, he emphasizes, takes place on the CPU core — no GPU required, which means essentially any Epyc customer can take advantage of the software platform.\n\nWhile no GPU may be required, it's worth noting that in virtualized environments, the software requires one core per VM or VPC, so the overhead scales with virtualization. Container runtimes presumably don't suffer from this same limitation. We've reached out to AMD for clarification.\n\n### 99 percent of the perf, half the DRAM\n\nAccording to AMD, the technology actually works. Initially, Smerdon says, the company targeted memory-hungry applications like electronic design automation (EDA) and content creation, but later expanded to support latency-sensitive applications like databases and [agentic AI](https://www.machinebrief.com/glossary/agentic-ai) pipelines.\n\n\"When pages of memory are cold, we can push them off to flash, and there's literally no impact on performance,\" he said.\n\nSmerdon is overstating things just a little here. Depending on the application and how much memory is offloaded to flash, there is a performance penalty.\n\nAccording to tests against Synopsys' PrimeTime EDA suite, he claims, half the memory can be offloaded to flash using Mext and the system still achieves 95 percent of the performance of DRAM. With 25 percent DRAM and 75 percent flash, performance fell to 80 percent of the DRAM-only baseline in AMD's tests.\n\nEDA tools are extremely memory-hungry, but the amount of data the cores are working on at any given moment is often measured in megabytes. This is why AMD's Epyc X-chips remain so popular with EDA customers and why Mext targeted it initially.\n\nIn other scenarios, like Apache Spark, AMD's Mext team says it is seeing 99 percent of the performance even after replacing half the DRAM in the system with flash.\n\nWhile implementing the tech may cost some performance, AMD is betting that customers will be willing to make that compromise in exchange for substantial DRAM savings, as memory remains constrained thanks to the AI datacenter buildout.\n\nRight now, AMD is working to integrate Mext into its broader enterprise portfolio. The software, we're told, will be exclusive to AMD hardware and will be offered as a license sold through OEM partners.\n\nSystems running Linux require only a kernel driver to take advantage of the software, while a separate RAM View application is being offered to help enterprises understand which systems and workloads it should be enabled on.\n\nThe chip designer says it's currently in the first phase of its Mext rollout, which means working with OEMs like HPE to ensure systems are so-called \"Mext ready.\" In the second phase, the company aims to begin selling licenses for the software and shipping Epyc platforms with RAM View preinstalled.®\n\nGet AI news in your inbox\n\nDaily digest of what matters in AI.", "url": "https://wpnews.pro/news/amd-where-we-re-going-we-don-t-need-as-much-dram", "canonical_source": "https://www.machinebrief.com/news/amd-where-were-going-we-dont-need-as-much-dram-autp", "published_at": "2026-11-11 09:59:00+00:00", "updated_at": "2026-11-11 10:19:44.356130+00:00", "lang": "en", "topics": ["ai-infrastructure", "machine-learning", "ai-chips", "artificial-intelligence"], "entities": ["AMD", "Mext", "Gary Smerdon", "Epyc", "Meta"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/amd-where-we-re-going-we-don-t-need-as-much-dram", "markdown": "https://wpnews.pro/news/amd-where-we-re-going-we-don-t-need-as-much-dram.md", "text": "https://wpnews.pro/news/amd-where-we-re-going-we-don-t-need-as-much-dram.txt", "jsonld": "https://wpnews.pro/news/amd-where-we-re-going-we-don-t-need-as-much-dram.jsonld"}}