{"slug": "amd-snaps-mext-to-break-the-memory-wall", "title": "AMD Snaps MEXT to Break the Memory Wall", "summary": "AMD acquired predictive memory technology startup MEXT to address the memory wall bottleneck in AI deployments. MEXT's AI-driven software optimizes memory by combining DRAM and flash, aiming to reduce server memory costs that now account for nearly 60% of server cost. The acquisition underscores the growing importance of memory management in data centers as AI workloads strain DRAM supply and costs.", "body_md": "The memory wall is real, and AMD’s acquisition of predictive memory technology specialist MEXT is a testament to the premise that AI scaling is increasingly a memory problem, not just a compute problem. AMD, which sells CPUs and GPUs for data centers and other AI deployments, is adding software-driven memory optimization to better manage memory access and efficiency in data center environments.\n\nThis acquisition highlights a growing emphasis on managing memory bottlenecks in cloud computing and AI deployments, and how improvements in memory management could potentially enable more efficient AI workload deployments and reduced total cost of ownership for data center operators. MEXT’s memory solution promises DRAM-class performance with flash-level economics and capacity.\n\nIt’s important to note that memory, not compute, is a constraint that keeps showing up in large AI deployments, and here DRAM has become the scarcest and most expensive part of the server design. “As DRAM supply remains constrained and memory costs stay elevated, enterprises are paying closer attention to software-driven approaches that can improve utilization and reduce overprovisioning,” said Mast Eastwood, senior VP of enterprise infrastructure at IDC.\n\n**AI’s memory crunch**\n\nLarge datasets used in AI models, data analytics, virtualization, and high-performance computing have made memory a central bottleneck. And memory is becoming a bottleneck in terms of bandwidth, latency, capacity, cost, utilization, and how efficiently workloads move data.\n\n[View All](https://www.eetimes.com/category/sponsored-content/)\n\nXbox CEO Asha Sharma recently made waves in trade press when she said that memory costs have risen roughly fivefold over the past two years. As a result, Xbox is unable to make as many consoles as consumers want.\n\nDRAM—unlike processors, networking, and storage—has remained architecturally unchanged since Intel introduced the first device in 1970. And every attempt to displace DRAM’s decades-old architecture has failed. As predicted, DRAM cost reductions and capacity growth stagnated while billions were invested in potential solutions, such as Intel Optane, but none succeeded.\n\nIn March 2023, Gary Smerdon founded MEXT to develop a predictive memory technology that can drastically reduce the cost of memory in servers while enabling much larger capacities. The upstart aimed to tackle the largest cost component in data centers: server memory or DRAM. Then, DRAM accounted for about 50% of server’s cost. Now, in 2026, with AI accelerating and DRAM demand surging, it accounts for nearly 60% of a server’s cost.\n\nMEXT’s solution was to create a hybrid architecture that combines DRAM and flash. However, while flash offers compelling economics, roughly 50× lower cost and 30× lower power than DRAM, the challenge was that flash, a storage technology, is 500× slower than DRAM. But what if cold memory pages—not actively used—are moved to flash and then AI algorithm predicts when they will be needed next?\n\nThat premise became the foundation of predictive memory technology, which targets infrastructure efficiency rather than model architecture. Consequently, small improvements in memory utilization can translate into meaningful gains in throughput, hardware utilization, and total cost of ownership at scale.\n\nMEXT’s solution: AI-driven memory optimization to tackle one of the biggest bottlenecks in modern compute infrastructure—the [memory wall](https://www.eetimes.com/the-memory-wall-is-real-here-is-the-door/)—which limits performance due to constraints in memory capacity and speed. But the Santa Clara, California-based company was set to tackle a problem many considered unsolvable: inefficiency in DRAM and a $100 billion market with chronic underutilization.\n\n**DRAM with flash economy**\n\nMEXT develops AI-powered software to facilitate inexpensive flash storage. Its memory optimization technology makes NAND flash behave like DRAM through AI-powered predictive memory, reducing infrastructure costs and expanding usable memory capacity without sacrificing performance.\n\nHere’s how this memory-tiering technology makes NAND flash appear as DRAM to the operating system; it works in three steps to create a new price-performance tier in which DRAM-class performance comes with flash-level economics and capacity.\n\nFirst, MEXT identifies memory pages that are not actively in use—termed as cold—and offloads them to flash, which costs 50× less than DRAM. Second, MEXT’s AI-powered engine predicts which offloaded memory pages are likely to be requested by the application soon. It continuously analyzes memory access patterns and uses AI models to anticipate which data stored in flash will be needed next.\n\nThird, the AI engine proactively pushes those pages back to DRAM before they are required. So, all relevant memory pages are always findable in DRAM, maintaining performance. As a result, application performance remains intact within a much smaller DRAM footprint, yielding substantially lower costs.\n\n**AMD’s end game**\n\nAMD’s acquisition of MEXT, first and foremost, unveils the company’s aim to tackle one of the most stubborn bottlenecks in modern AI computing: memory. MEXT’s predictive AI models reduce latency and optimize data pathways between CPUs, GPUs, and [high-bandwidth memory (HBM)](https://www.eetimes.com/hbm-innovation-outpaces-standards-development/), expanding usable memory capacity without forcing data center operators to proportionally increase their hardware spend.\n\nAMD’s MI300 and MI355X accelerators are heavily reliant on HBM, so more efficient memory utilization translates directly into superior performance-per-watt. However, integrating MEXT’s predictive memory technology with AMD’s CPUs and AI accelerators will require effective execution. A lot depends on how intelligently AMD’s AI systems manage memory, data movement, and workload behavior at scale.\n\nThe acquisition also marks an aggressive push to reshape the company from a traditional CPU and GPU vendor into a full-stack AI powerhouse capable of competing with Nvidia in both hardware and the critical software ecosystem. AMD is strengthening its [ROCm](https://www.eetimes.com/taking-on-cuda-with-rocm-one-step-after-another/) software ecosystem against Nvidia’s CUDA, and the NEXT toolchain could help the company close the gap with Nvidia’s formidable CUDA channel. It could also help AMD to become a legitimate second source to Nvidia in the AI accelerator market.\\AMD’s acquisition of MEXT is an important landmark in the industry’s quest to break the memory wall for both data center CPUs and AI accelerators. It’s also an important strategic step for AMD to take a more holistic look at AI system designs.\n\n##### See also:\n\n[AI-Driven Memory Shortage Upends IT Budgets](https://www.eetimes.com/ai-driven-memory-shortage-upends-it-budgets/)\n\n[Massive AI Storage Demand Creates a New Memory Wall](https://www.eetimes.com/massive-ai-storage-demand-creates-a-new-memory-wall/)\n\n[Alliance Aims to Deliver Memory-Optimized AI for Inferencing](https://www.eetimes.com/alliance-aims-to-deliver-memory-optimized-ai-for-inferencing/)", "url": "https://wpnews.pro/news/amd-snaps-mext-to-break-the-memory-wall", "canonical_source": "https://www.eetimes.com/amd-snaps-mext-to-break-the-memory-wall/", "published_at": "2026-06-16 13:50:23+00:00", "updated_at": "2026-06-16 13:56:35.102070+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-infrastructure", "ai-chips", "ai-startups", "ai-tools"], "entities": ["AMD", "MEXT", "Intel", "IDC", "Xbox", "Asha Sharma", "Gary Smerdon", "NAND"], "alternates": {"html": "https://wpnews.pro/news/amd-snaps-mext-to-break-the-memory-wall", "markdown": "https://wpnews.pro/news/amd-snaps-mext-to-break-the-memory-wall.md", "text": "https://wpnews.pro/news/amd-snaps-mext-to-break-the-memory-wall.txt", "jsonld": "https://wpnews.pro/news/amd-snaps-mext-to-break-the-memory-wall.jsonld"}}