Is the GPU not the real bottleneck of AI? Intel bets everything on memory

With the evolution of AI-based workloads, the race for performance is no longer just about GPUs and dedicated accelerators. Intel believes that one of the main limiting factors in modern datacenters is now the platform's ability to quickly transfer data between processors, memory, and accelerators.

For this reason, the company has announced support for DDR5 RDIMM memory at 8000 MT/s on the Xeon 6 platform, becoming the first manufacturer to offer this speed in the mainstream server segment. The upgrade will affect specific versions of the Xeon 6700P "Granite Rapids" processors and Xeon 6+ "Clearwater Forest," which will receive support through a UPLR update and BIOS update. Production availability is expected between August and September 2026.

According to Intel, the transition from the current RDIMM DDR5 modules at 6400 MT/s to the new solutions at 8000 MT/s results in a 25% increase in memory speed, accompanied by a latency reduction of about 6%. Overall, the platform would be able to offer up to 20% additional bandwidth, a crucial element in AI systems, where host CPUs, GPUs, storage, and networks must operate together without creating bottlenecks.

The company emphasizes that modern datacenter CPUs are called upon to manage increasingly larger artificial intelligence models, configurations with a growing number of accelerators, and processing pipelines characterized by a high volume of data. In this context, memory is no longer just a technical feature but becomes a determining element to keep CPUs and GPUs powered without them being idle waiting for data transfer.

In internal tests, Intel indicates performance improvements ranging from 3% to 6% across a broad set of applications, with more pronounced benefits in workloads particularly sensitive to bandwidth. The goal is to create more balanced platforms, where the increase in the number of cores or the power of accelerators can translate into an actual increase in overall performance.

The announcement also includes a roadmap dedicated to second-generation MRDIMM (Multiplexed Rank DIMM). An MRDIMM is a memory module designed to increase data transfer speeds in servers without altering the format of traditional RDIMMs. It utilizes a small chip called a multiplexer (mux), which allows simultaneous access to multiple memory ranks present on the module instead of alternating access. The result is higher throughput, useful for applications such as artificial intelligence, scientific simulations, and data-intensive analyses.

Intel plans to enable modules up to 8800 MT/s on the Xeon 6900P "Granite Rapids" processors by the first quarter of 2027. This solution will be primarily aimed at systems with a high number of cores, from 72 to 128, employed in artificial intelligence, data analysis, and high-performance computing (HPC) applications.

The outlook is for continued growth in memory speeds. According to the illustrated roadmap, future generations of MRDIMM could reach 12,800 MT/s as early as 2028 and go up to 17,600 MT/s between 2029 and 2030. An evolution that, according to the company's intentions, would enable DDR5 memory to remain competitive even during the transition to future DDR6 memories.