On November 18 (US time; November 19 in Japan), JEDEC announced its development plan for next-generation DDR6 DRAM. For server registered DIMMs (RDIMMs), the starting point is 12.8 GT/s, with 14.4 GT/s and higher envisioned beyond that. The proposal also increases the number of data signal lines from DDR5's 64 to 96.

The same 12.8 GT/s is already achieved by the DDR5 MRDIMM supported by AMD's EPYC 9006 server CPUs. A DDR5 MRDIMM collects data from multiple DRAMs on the module and sends it in sequence over the same lines. The DDR6 proposal instead widens the set of data lines that can be used in parallel, so more data moves in each transfer.

AD

Same 12.8 GT/s, different data per second

JEDEC's DDR6 preview announcement proposes increasing the number of DQ signal lines, which carry data, from 64 in DDR5 to 96. GT/s is the number of transfers per second on each signal line, while bandwidth in GB/s shows how much data all the lines together carry per second. So even if the per-line transfer rate is equal, adding lines that work in parallel raises theoretical bandwidth.

According to AMD's technical paper on EPYC 9006, DDR5 MRDIMM Gen2 runs its individual DRAMs at about 6,400 MT/s while sending data to the CPU at 12,800 MT/s. It does this through "rank multiplexing": two ranks, meaning sets of DRAMs that are read and written together, operate in parallel, and a buffer places their data in turn onto the faster host-side bus. Since 1 GT/s equals 1,000 MT/s, the transfer rate seen from the host matches DDR6's initial target.

At the same 12.8 GT/s, the 96-line DDR6 proposal gives 153.6 GB/s, versus 102.4 GB/s for a 64-line DDR5 MRDIMM equivalent.

Comparison Host-side transfer rate Data lines Theoretical peak bandwidth
JEDEC DDR6 RDIMM initial proposal 12.8 GT/s 96 153.6 GB/s
AMD EPYC 9006 DDR5 MRDIMM Gen2 12,800 MT/s 64 (equivalent) 102.4 GB/s

Bandwidth is calculated as transfer rate (GT/s) × number of data lines ÷ 8. The comparison covers the theoretical upper limit of the data bus, excluding signals used for ECC. The DDR5 figure matches the per-channel peak bandwidth given in Section 3.1 of AMD's paper.

Widening the data bus by 1.5 times at the same speed raises theoretical peak bandwidth by 1.5 times. In its preview, JEDEC cited 9.2 GT/s as the comparison figure for DDR5 RDIMM transfer rates and set a goal of raising bandwidth by more than 50% when DDR6 is introduced. It has not published the detailed comparison conditions behind that launch target.

MRDIMM gains bandwidth without raising the DRAM's own transfer rate to the host-side level. The module needs circuitry to gather and send out the data. AMD explains that, because of the faster host side and this added circuitry, module power consumption is somewhat higher than for comparable RDIMMs.

Banks and refresh to keep the bus busy

DDR6 plans to double the number of banks and improve refresh. A bank is a unit within a DRAM that allows accesses to be separated, and refresh is the operation that periodically restores the charge holding data. The target bank cannot be accessed during refresh, so even with a faster, wider external bus, waits occur if data cannot be supplied from inside.

DDR5 also has measures to reduce those waits. Micron's DDR5 technical explainer describes a configuration that divides the DIMM into sub-channels, along with same-bank refresh. The latter is a command targeting banks with the same number in each bank group, reducing the need to take all banks offline at once and allowing access to other banks to continue.

DDR6 also plans to divide access into "pseudo-channels" to increase bandwidth per package and configuration flexibility. With more banks, there are more options for steering accesses away from a bank being refreshed toward another one. How much idle bus time this removes depends on the distribution of accessed locations and on memory controller behavior. The preview does not disclose the number or width of pseudo-channels, or the specific refresh procedures.

AMD's paper cites CPU-based AI inference and sparse matrix computation as examples where main memory bandwidth tends to limit processing. In GPU-based systems too, if main memory access becomes congested during data preprocessing handled by the CPU, data delivery to the GPU is delayed. In such cases, increasing main memory bandwidth is likely to show an effect. How much an application as a whole speeds up depends on how much of its processing time is spent waiting for data from main memory.

AD

Higher density in 74 mm

JEDEC showed a module length of 74 mm, versus 133 mm for DDR5 as the comparison. On capacity, it set out support for up to 128 Gb per DRAM die and plans to optimize die stacking. Dividing 128 Gb by 8 gives the equivalent of 16 GB per die. Module capacity depends on the number of packages and the number of dies in each package.

Increasing die capacity and stacking dies in the same package makes it easier to raise the capacity that fits on a given board area. A shorter module could also free up space for socket placement on server boards. Placing many sockets in a small area requires fitting the routing for 96 data lines and securing room for power, cooling, and hands-on access for replacement.

JEDEC listed RDIMM for servers, unbuffered UDIMM for ordinary desktop PCs, and CAMM, which is pressed against and fixed to the board, as candidate form factors. The 74 mm figure is the dimension shown in the preview; dimensions and terminal details for each form factor have not yet been published.

Power, reliability, and what to look for in products

DDR6's power measures include I/O circuitry terminated on the ground side (VSS) and low-power modes for data centers. Termination suppresses signal reflections at the end of a transmission line and is part of the electrical design that makes high-speed I/O work. JEDEC aims to improve performance per watt. The cooling and power supply capacity needed when adopting the memory is determined by the power that the memory controller and modules consume under load.

On reliability, JEDEC envisions multiple implementations of system-level error correction (ECC). The on-die ECC for DDR5, as described in Micron's paper, corrects errors in data stored inside the chip before it is output. System-level ECC handles the data the memory system receives. In DDR6 products, the range of errors that can be corrected, and the wiring and power needed for that, will shape designs.

The LPDDR6 standard JESD209-6, released by JEDEC in July 2025, is a separate standard from this DDR6. The current DDR6 proposal could change during the approval process. The release date of the finalized standard and the launch dates of compatible CPUs and memory products have not been given.

To gauge the benefits of a generational change in servers, you need to look at the bandwidth a CPU can actually use under real workloads, together with power consumption including the controller and modules. For desktop PCs, 12.8 GT/s is the initial target for server RDIMMs, so comparisons with existing machines will only be possible once the speeds and capacities of compatible CPUs and UDIMMs are known.