Server ECC RDIMM vs DDR5 (2026): 4x Cheaper per GB

Bottom Line

• DDR4 ECC RDIMM: $9.14/GB median. Range $4.53 to $14.34/GB. Cheapest listing is an A-Tech 16GB DDR4-2133 module at $4.53/GB.
• DDR5 ECC RDIMM: $38.48/GB median. Range $29.56 to $46.49/GB. That is a 4.2x gap.
• 512GB costs roughly $4,680 in DDR4 at the median, or about $19,700 in DDR5. For capacity-bound CPU inference, this is not a close decision.
• The honest catch: DDR4 is not cheap, it is only cheaper. It has risen 60-80% since early 2025. A 4x discount on an inflated market still costs thousands.
• RDIMM does not fit a desktop board. You are buying a server platform too. Budget for it.
• Buy channels, not speed. Eight populated channels beat four fast modules for inference throughput.

All prices read from the DatacenterDisk live tracker on 2026-09-07. That tracker refreshes every four to five hours and the DDR5 median moved 3.9% in the week before we read it. Treat any figure here as a snapshot, not a quote.

Ready to buy? See the tested hardware list with current prices.

Why This Question Exists At All

Large mixture-of-experts models changed the arithmetic of local inference. A model with 400B total parameters but 20B active per token does not need 400GB of fast VRAM. It needs 400GB of some memory and enough bandwidth to stream the active experts.

That reframes the build. Instead of chasing an impossible stack of GPUs, you buy a lot of ordinary system RAM and run on CPU. We walked through that argument in EPYC server CPUs for MoE inference.

The moment you accept that framing, one question dominates the budget: what is the cheapest way to buy 256GB or more of RAM?

The Gap, Measured

Read 2026-09-07 from the DatacenterDisk live tracker.

The ratio at the median is 4.2x. At the cheapest-listing end it is 6.5x, because the DDR4 market has a long tail of old slow modules and DDR5 does not yet.

What that buys you, at 2026-09-07 prices:

For a build whose entire purpose is holding a large model in memory, that table ends the argument. You go DDR4.

The Three Things That Discount Costs You

Most pages stop at the ratio. The ratio is the easy half.

1. RDIMM will not go in your desktop

Registered DIMMs have a register between the memory controller and the DRAM chips. It is what lets a board address this much memory at all, and it is why a consumer AM5 or LGA1700 motherboard physically will not run them.

So the real comparison is not “DDR4 RDIMM versus DDR5 UDIMM”. It is “a used server platform plus DDR4 RDIMM” versus “a desktop plus DDR5 UDIMM”. Add the board and the CPU before you compare.

For single-socket SP3 EPYC builds we point at the ASRock Rack ROMED8-2T, because it gives eight DDR4 channels, which is the specification the whole MoE argument rests on.

2. Cheap DDR4 is cheap because it is slow

That $4.53/GB listing is DDR4-2133. It is the oldest speed grade in the tracker, and that is exactly why it is at the bottom of the range.

Once the model fits in memory, bandwidth sets your tokens per second, not capacity. A build that fills all eight channels with DDR4-3200 will generate meaningfully faster than the same capacity in four channels of anything.

The practical rule: buy the number of modules that fills every channel on your board first, then optimise speed grade with whatever budget is left. Eight slow modules beat four fast ones.

3. “Four times cheaper” is not “cheap”

This is the sentence the buying guides skip.

DDR4 server memory has risen 60-80% since early 2025. DDR5 rose 100-116%. Both markets are casualties of the same thing: wafer capacity reallocated toward HBM for AI accelerators, which is the memory that goes into the datacenter GPUs everyone is buying.

So DDR4’s advantage is real but it is a discount on a market that already moved against you. 512GB at $4,680 is the cheap option in 2026. Two years ago the same capacity cost a fraction of that. If you are comparing against a price you remember, you will be unpleasantly surprised — the same trap we documented for consumer memory in should I buy RAM now?.

When DDR5 Is Still Right

Three cases, and none of them is “it is newer”.

• You already own the platform. If you have a current-generation server board, the RDIMM type is decided for you.
• Your capacity target is small. At 64GB or 128GB the absolute gap is a few thousand dollars rather than fifteen, and DDR5’s per-channel bandwidth advantage starts to matter more than the price.
• The box is not only for inference. If it runs databases or builds as well, buying current-generation hardware you can keep for six years is a defensible call.

For a machine bought specifically to hold a large MoE model, none of these apply.

What We Would Actually Build

If capacity is the goal and the budget is real: a single-socket SP3 EPYC board with all eight DDR4 channels populated, using 32GB or 64GB registered ECC modules rather than the cheapest 16GB ones. That lands 256GB to 512GB somewhere between $2,300 and $4,700 in memory, on a platform that costs a fraction of a current-generation equivalent.

Then check the bandwidth you actually got before you tune anything else. Half the disappointing CPU-inference builds we see are correctly sized and wrongly populated — the RAM is all there, sitting on two channels.

See Also

• EPYC server CPUs for MoE inference — the platform this memory goes into, and why eight channels is the number
• Should I buy RAM now? The 2026 shortage — the same market forces on the consumer DDR5 side
• Best local LLM for 128GB VRAM — the GPU path this build is the alternative to
• Cheapest 32GB VRAM GPU in 2026 — if you would rather buy VRAM than system RAM
• Best local LLM for 16GB VRAM — the other end of the budget, where capacity is the binding constraint too