AMD acquires MEXT to revolutionize data center memory

What happened?

AMD has announced the acquisition of MEXT, a startup that develops a Predictive Memory Engine that moves infrequently accessed data from DRAM to NAND storage, making flash appear as main memory to applications. This memory tiering technology directly addresses memory constraints in data centers, where DRAM is expensive and scarce. According to Tom's Hardware, the acquisition was confirmed by AMD, though financial terms were not disclosed. Speculation suggests the deal could be around $100 million, given MEXT's size and early-stage technology. Founded in 2020, MEXT had raised approximately $20 million in previous funding rounds, according to Crunchbase data.

MEXT's predictive engine uses machine learning algorithms to identify data access patterns and predict which data will be needed soon. This keeps hot data (frequently used) in DRAM and migrates cold data to NAND, which is up to 10 times cheaper per gigabyte. NAND flash latency is orders of magnitude higher than DRAM (microseconds vs nanoseconds), but the predictive engine minimizes impact by anticipating needs. In internal tests, MEXT reported up to a 40% reduction in total cost of ownership (TCO) for in-memory database workloads, according to leaked technical documents.

Why is this important?

Memory demand in data centers is growing exponentially with workloads like AI, real-time analytics, and in-memory databases. However, DRAM has physical and cost limitations: DRAM prices have fluctuated historically, but in 2023, the price per gigabyte of DRAM was roughly 10 times higher than NAND flash (according to TrendForce). Additionally, DRAM scalability is limited by module capacity and server slot count. MEXT's solution allows applications to benefit from much larger memory capacity (using flash) without sacrificing performance, as the predictive engine keeps hot data in DRAM.

This acquisition strengthens AMD's position against competitors like Intel (with its Optane technology, discontinued in 2022) and NVIDIA (with its Grace Hopper platform integrating unified memory). Intel had bet on Optane as a persistent memory solution, but its high cost and complexity led to its demise. Now, AMD seeks to fill that gap with a more efficient and cost-effective solution. It also aligns with AMD's strategy to offer comprehensive data center solutions, from EPYC CPUs to Instinct GPUs and now intelligent memory management. In the server market, AMD has gained share against Intel, rising from 10% in 2020 to over 30% in 2023, according to Mercury Research. This acquisition could accelerate that trend by offering a differentiating memory advantage.

Historically, memory tiering has been explored by startups like MemVerge, which combines persistent memory with management software, and open-source projects like TPP (Transparent Page Placement) in Linux. However, MEXT differentiates itself with its ML-based predictive approach, promising dynamic adaptation to workloads. Integration with AMD's EPYC platforms will enable hardware-level optimization, something purely software solutions do not offer.

Market implications

• For enterprises: Cost reduction by using cheaper flash storage instead of additional DRAM. For example, a company running in-memory databases like SAP HANA could cut its memory budget by 30-50%, according to analyst estimates. Additionally, greater effective memory capacity will allow handling larger datasets without horizontal scaling, simplifying infrastructure.
• For users: Faster applications with greater in-memory data capacity, especially in cloud and big data environments. End users will experience lower latency in AI and analytics applications, though the impact will be indirect through service providers.
• For the industry: Accelerated adoption of memory tiering as a data center standard. Other chipmakers like Intel and NVIDIA are expected to develop similar solutions or acquire competing startups like ScaleMP or RamCache. Possible consolidation of similar startups: companies like Pliops, which accelerates storage with storage processors, could be acquisition targets. It could also boost investment in persistent memory technologies like MRAM or ReRAM, though these are still early-stage.

What readers should know

MEXT's technology is not immediately available; AMD is expected to integrate it into future EPYC platforms, likely the next generation (Turin) based on Zen 5 architecture, slated for 2024-2025. Financial details of the acquisition have not been disclosed. Speculation: it could be a deal under $100 million, given MEXT's size and lack of significant revenue. However, the strategic value could be higher if the technology becomes a key differentiator. AMD has not confirmed whether the technology will be offered as part of its CPUs or as a standalone chip (similar to an FPGA or ASIC).

"Memory is the new bottleneck in data centers. With MEXT, AMD offers an intelligent solution that maximizes performance per dollar." — Analyst at TheVortiq

In summary, this acquisition is a strategic move that strengthens AMD's roadmap in high-performance computing and efficiency. The combination of EPYC CPUs, Instinct GPUs, and now predictive memory management positions AMD as a comprehensive data center infrastructure provider, directly competing with NVIDIA's integrated solutions and Intel's future offerings. However, questions remain about the technology's maturity and practical integration. Readers should closely follow AMD's announcements on future platforms and independent benchmarks that validate performance claims.