JEDEC has published an early draft of DDR6. This is a major shift in memory architecture for the future. This is important because the new standard is going to offer more than 50% bandwidth increase at release compared to DDR5.
JEDEC outlines early DDR6 specifications including a wider bus and smaller modules
This is going to have a big impact on performance of data driven applications. The changes aren't just incremental, but a different way of how physical and logical architecture of DRAM modules are going to be.
DDR6 Specifications
- Data Bus Width: 96 DQ
- Initial Data Rate: 12.8 GT/s
- Module Length: 74 mm
- Max DRAM Die Density: 128 Gb (16 GB)
- Architecture: Pseudo-channel, doubled bank count
A major part of the new standard is a huge increase in the data bus width. JEDEC is going to expand the bus width from 64 DQ to 96 DQ. That's a 50% increase in the bus width.
This increase is going to provide more room to transfer more data simultaneously between memory and CPU. This is going to be the basis for the claimed speed increase.
There are a few more changes in the architecture for DDR6. These are aimed to improve the efficiency and the density. Pseudo channel architecture and a double amount of banks are going to be used to manage the data flow more efficiently.
JEDEC is going to support up to 128 Gb of DRAM dies. (16 GB per die). This will enable more capacity for modules.
DDR6 modules will have a size of 74mm. That's a major decrease in size compared to DDR5 (133mm). The initial data rate will be 12.8 GT/s.
This is an increase compared to current DDR5 (9.2 GT/s). There are plans for 14.4 GT/s and more.
These are not final specifications. DDR6 standard is currently in development. JEDEC mentions that proposed features and specifications are subject to change during development and approval.
This early draft gives a good direction of where things are heading. But final implementation depends on the final standard.
Source: NotebookCheck

The jump from 64 to 96 DQ is the real story here. Doubling the bank count alongside that wider bus should finally stop memory from being the bottleneck for multi-core workloads. Pseudo-channel architecture is a smart move to keep signal integrity manageable at those speeds.