JEDEC Reveals DDR6 Details: 50% More Bandwidth, Higher Capacity, Shorter DIMMs

Jedec has published its first details on DDR6, the next generation of DRAM, and the standards body is committing to at least 50 percent more bandwidth than DDR5, higher capacities per module and a shorter DIMM — though the specification is still in development and no completion date has been given.

The details come from a Jedec press release, Tweakers reported on Wednesday.

The headline numbers are substantial. DDR6 starts at 12.8 gigatransfers per second, against DDR5’s official ceiling of 9.2GT/s, and Jedec already has speeds of 14.4GT/s and beyond on its roadmap, the standards organisation said.

The memory bus widens too, from 64-bit to 96-bit. Both changes push raw bandwidth up, which is the metric that matters most to the buyers Jedec is actually aiming at first: datacenter operators feeding AI accelerators and large-memory servers.

Capacity is the other pillar. DDR6 modules will support DRAM chips of 128Gbit each — 16GB per chip — and Jedec says it will become possible to stack multiple DRAM chips on top of one another, raising capacity further.

That stacking may prove necessary, because the new form factor shrinks the module itself: DDR6 DIMMs measure 74mm long, down from 133mm for DDR5. Shorter modules with stacked chips mean more memory per unit of motherboard real estate, which is exactly the trade server designers care about.

What buyers do not get is compatibility. As with every previous DDR generation, the new modules will not work in existing DDR5 motherboards. A DDR6 platform means a new board, a new CPU support cycle and a new procurement line item.

The spec sheet, side by side

Specification DDR5 DDR6
Data rate at launch Up to 9.2GT/s From 12.8GT/s, 14.4GT/s and higher on the roadmap
Bus width 64-bit 96-bit
DRAM chip density — 128Gbit (16GB) per chip
Module length 133mm 74mm
Chip stacking — Supported

Jedec is also promising better power efficiency than DDR5, and it is splitting memory access into what it calls pseudochannels. The organisation shared few details on that mechanism, saying only that moving data with pseudochannels happens more efficiently and that this should lift bandwidth further.

Readers who remember DDR5’s own channel-splitting architecture will recognise the pattern: finer-grained access is how DRAM standards have been buying bandwidth without relying on raw clock speed alone.

Why datacenters get it first

Consumers will wait. Jedec has not said when the standard will be finished, and in its first incarnation DDR6 memory is aimed mainly at datacenters.

AMD said its Ryzen processors will support DDR6 when that really makes sense, and Intel likewise appears in no hurry to move its desktop or laptop CPUs off DDR5. That leaves the client market on DDR5 for years — hardly unusual, given that DDR5 itself took several years after its 2020 standardisation to reach mainstream pricing.

There is a live argument about whether mainstream desktops ever need it. On Tweakers, commenter Scribe argued that LPDDR is closing the gap: LPDDR5X is now faster and more power-efficient than DDR5, and with the modular SOCAMM2 form factor, soldered-down memory’s main drawback is fading. On that reading, DDR6 could end up a server-only standard focused on extreme capacity per square centimetre.

Tweakers reader klakkie.57th argued that DDR6 is only necessary if one wants 10+ TB on a single CPU, adding that with current memory prices, it would be better to add extra CPUs.

That comment, from Tweakers reader klakkie.57th, cuts at the economics. But the same reader then answered the objection himself, pointing to NUMA — non-uniform memory access — the design in multi-CPU systems where a processor reaches its own locally attached memory faster than memory attached to another processor.

A SQL Server instance on one 64-core CPU with 2TB of memory, he wrote, works better and faster than the same load split across two 32-core CPUs each addressing 1TB. One large memory pool per socket is precisely the problem high-capacity, high-bandwidth DDR6 is built to serve.

The 360-company bottleneck

How long until shipping silicon? Nobody has said, and the standardisation process itself is part of the answer. Jedec’s membership runs to roughly 360 companies, as one Tweakers commenter noted in the discussion under the report — every one of them with a stake in the approval procedure.

Another reader called that bureaucracy a problem, while a counterpoint in the same thread held that slow, broad consensus is arguably a blessing: it keeps the platform stable enough to build on.

History supports both readings. Each DDR generation has taken years from final specification to affordable consumer modules, and DDR6’s datacenter-first targeting suggests the client transition will be slower still. Meanwhile the competing pressure is real: LPDDR variants keep improving on speed and efficiency, and Jedec’s own SOCAMM2 work gives low-power memory a modular path into servers. The gap DDR6 must justify is narrower than the one DDR5 faced.

What enterprise buyers should watch

For now, nothing changes in the rack. The specification is unfinished, the DIMM geometry is new, and neither AMD nor Intel has committed client or server timelines.

The practical signals to watch are the final Jedec standard’s publication, the first CPU platform announcing DDR6 support, and whether the promised pseudochannel architecture survives contact with real controllers. Until then, DDR5 remains the shipping standard — officially up to 9.2GT/s, on a 64-bit bus, in a 133mm module.

Photo of author

Technology Desk

Technology Desk at Archyde.

Nina Pinzarrone Leaves US Training Camp Over Unacceptable Behavior