Memory bandwidth, not core count, is the quiet limiter on a growing share of server workloads - and the roadmap to fix it is now visible. DDR5 is mature, MRDIMM is shipping on the latest platforms as a bandwidth booster, and DDR6 is the next full generation on the horizon. For UK buyers planning a refresh that has to last four or five years, the timing of these transitions matters as much as the speeds: buy too early into a young technology and you pay a premium for thin support; wait too long and you leave bandwidth on the table. Here is how to plan around the memory roadmap rather than be surprised by it.
Why memory bandwidth is the new bottleneck
Core counts have raced ahead faster than the memory subsystem that feeds them. Pack 128 or more cores into a socket and they contend for the same memory channels, so per-core bandwidth falls even as total compute rises. For analytics, large in-memory databases, dense virtualisation and the data-loading side of AI, the host stalls waiting for memory long before it runs out of cores. That is why bandwidth, not capacity or clock, increasingly decides real throughput.
The industry's answer comes on two tracks: widen and speed up the channels (more channels, then DDR6), and lift effective bandwidth on the existing DDR5 channels with MRDIMM. UK buyers do not have to pick one - the roadmap layers them - but you do need to know which is available when, so you can time a purchase to the workload's appetite.
MRDIMM: more bandwidth on DDR5 platforms now
MRDIMM (multiplexed-rank DIMM) is the near-term bandwidth lever. It uses a buffer to present multiple ranks in a way that lets the host fetch more data per cycle, raising effective bandwidth meaningfully over a standard DDR5 RDIMM at the same nominal generation. Crucially it runs on current platforms that support it - notably the latest Intel Xeon 6 - so it is available to buy today rather than being a future promise.
The catch is the usual early-technology one: MRDIMM modules command a price premium over standard RDIMMs and are supported only on specific platforms, so the value depends entirely on whether your workload is bandwidth-bound. For a memory-bound analytics or HPC-style host the uplift can justify the premium outright; for a general business server that never saturates its channels, standard DDR5 RDIMM remains the sensible, cheaper choice. Match the module to the workload using our memory and RAM guidance.
- •DDR5 RDIMM: mature, lowest cost per GB - the default for general workloads
- •MRDIMM: higher effective bandwidth on supporting platforms now, at a price premium - for bandwidth-bound hosts
- •DDR6: the next full generation, a larger bandwidth and efficiency step - a future platform transition
DDR6: the next generation, and when it lands
DDR6 is the next full memory generation and promises a substantial step in both bandwidth and energy efficiency over DDR5, as every generation has. But a memory generation is a platform transition, not a drop-in upgrade: it requires new CPUs, new boards and a mature module supply, so it arrives gradually and at a premium before settling into the mainstream. Treat DDR6 as a roadmap item to design around, not a product to wait for if you need a server this year.
The familiar early-adoption curve applies: the first DDR6 platforms will carry a price premium, limited module choice and the teething issues of any new generation, then volume drives cost down and maturity up over the following couple of years. Plan to adopt DDR6 when the platform you want ships it as standard and supply is healthy - not on day one unless a specific workload demands it.
Planning a refresh across the transition
The practical framing is to map your refresh against the curve. If you must buy now, DDR5 is the safe, mature foundation, with MRDIMM as a targeted bandwidth upgrade only on the hosts that are genuinely bandwidth-bound. If a workload is on the edge of needing more bandwidth and your timeline is flexible, it can be worth aligning the purchase with platforms that have a clear path to MRDIMM or the next generation, so you are buying into rising rather than ending support.
Pair the memory decision with the platform decision - the CPU generation dictates what memory you can use, which channels you get and what is on the upgrade path. Choose the processor with our processors guidance, then size and balance memory in our server configuration service so every channel is populated and no bandwidth is stranded.
A buyer's rule of thumb
Keep it simple. For general-purpose servers, buy mature DDR5 RDIMM and do not pay for bandwidth you will not use. For demonstrably bandwidth-bound workloads on a supporting platform, MRDIMM is worth the premium now. For long-life strategic platforms, favour a CPU generation with a credible path to MRDIMM and DDR6 so the investment ages well. And in every case, populate and balance every channel - the most common, most expensive memory mistake is leaving bandwidth on the floor through a lopsided layout, regardless of which technology you bought.