JEDEC, the organization that sets the open standards for computer memory, has provided a new preview of DDR6. The transfer speed is heading toward 12.8 GT/s, or 12,800 MT/s. That’s about three times faster than the JEDEC standard for DDR5 at 4,800 MT/s. However, the more important specification is likely to be the one that makes the biggest difference: the data bus will be 50 percent wider, which makes a huge difference for data-hungry workloads such as AI. A DDR6 module will also be smaller than previous generations of system memory.
This preview comes just six years after the first DDR5 specifications were released. While JEDEC is outlining a successor, it will take at least another year before actual production units are available.
It is important, however, that the expected speed does not become merely a minimum. For example, DDR5 officially starts at 4,800 MT/s, but more common speeds today range from 5,200 to as high as 9,200 MT/s. In any case, the roadmap for DDR6 already lists 14.4 GT/s “and beyond.”
Speed is often not the most important factor for system memory; within a fairly wide bandwidth range, it is many times faster than SSD storage and much slower than HBM memory on GPUs or the cache within a CPU.
However, the data bus is expanding from 64 to 96 DQ, which alleviates a significant bottleneck. Combined with the speed gains, this should result in more than 50 percent additional bandwidth compared to DDR5 upon its introduction. JEDEC does emphasize, however, that this does not automatically translate to a corresponding improvement in application performance. This is because system memory itself is often not the bottleneck; rather, it is, for example, the operating system’s hardware scheduling or software design.
Pseudo-channels and shorter modules
A lot is also changing under the hood. DDR6 divides memory access into “pseudo-channels,” which should enable more bandwidth per DRAM package and move even further away from the classic dual/quad-channel design in which CPUs accept two or four separate data streams from DIMMs. In addition, the number of banks doubles, and the refresh rate will be faster. The maximum die density also increases: from 64 Gb in DDR5 to 128 Gb, with optimizations for chip stacking, also known as “die stacking.”
For the first time, the modules will be significantly smaller. A DDR6 module will be 74 mm long, compared to 133 mm for DDR5, which should enable higher memory density per system. Furthermore, DDR6 supports multiple ECC implementations, and an improved I/O mechanism is expected to significantly boost performance per watt. RDIMM, UDIMM, and CAMM are all being considered as form factors.
“As AI and data-intensive workloads continue to drive up system requirements, DDR6 is designed to deliver meaningful advances in memory bandwidth, capacity, and reliability,” says Christopher Cox, chair of JEDEC’s JC-42 committee.
When will DDR6 be available?
DDR6 is still some time away. JEDEC made an initial concept version available to manufacturers as early as late 2024. This year, Samsung, SK hynix, and Micron have begun early development, including prototypes with substrate suppliers. According to TrendForce, the first commercial products aren’t expected until 2028–2029, with servers and AI infrastructure as the initial target market. Consumer PCs will follow later and will initially be available only at high prices, assuming that prices even higher than today’s are affordable.
Read also: Memory chip prices skyrocket, with no end in sight