Data rate is not delivered bandwidth
Channel count, bus width, protocol overhead, access patterns, controller scheduling, and workload locality determine how much theoretical bandwidth becomes useful.
DDR5 is a deployed memory ecosystem. Desktop and server DDR6 are still an emerging standards and platform transition. The useful comparison is therefore between what exists now and the direction under development—not between two interchangeable products.
Scope: this page compares standard desktop/server DDR5 with the reported direction of standard desktop/server DDR6. Published LPDDR6 is a separate low-power standard and is covered on the LPDDR6 page.
Confirmed labels refer to established DDR5 characteristics or published facts. Reported labels identify DDR6 targets and architecture directions that should not be read as frozen requirements.
| Topic | DDR5 now | DDR6 direction | Confidence |
|---|---|---|---|
| Standards state | Published JEDEC generation with mature controller, DIMM, validation, firmware, and platform support. | Desktop/server generation moving through standards work, vendor development, validation planning, and platform roadmaps. | Reported direction |
| Data-rate range | JEDEC DDR5 began at 4,800 MT/s; later standard revisions extend the generation above its launch rate. Actual support depends on CPU and module class. | Industry reporting commonly discusses ranges beginning around 8,800 MT/s and extending materially higher over the generation. | Expected, not final |
| Channel organization | A conventional DIMM exposes two independent 32-bit sub-channels, excluding ECC bits. | A 4 × 24-bit organization has been reported as an architecture direction, potentially increasing independent transaction paths. | Reported architecture |
| Capacity & density | Multiple die densities and DIMM capacities ship across client and server markets. | Higher density is expected over the generation, but final device organizations and supported module capacities require standards and platform confirmation. | Platform dependent |
| Latency | Known timings vary by speed bin, module, controller, topology, and workload. | No responsible single latency number exists yet. More transfers per second do not guarantee lower access latency. | Not final |
| Voltage & power | Defined operating requirements and established module power-management implementations. | Efficiency per transferred bit is a design goal. Final desktop/server voltage and power requirements should not be inferred from rumors or LPDDR6. | Not stated |
| Module ecosystem | UDIMM, SODIMM, RDIMM and newer form factors serve established client and server use cases. | Module and attached-memory approaches may evolve. The eventual mix of DIMM, CAMM, SOCAMM, or other implementations is use-case and platform specific. | Roadmap dependent |
| Platform support | Available across shipping CPU families, chipsets, boards, firmware, test tools, and operating environments. | Requires a new compatible memory controller and PHY, electrical validation, board or package design, firmware, and qualified memory. | Confirmed principle |
| Availability | Shipping broadly in PCs, workstations, and servers. | No established mainstream desktop/server DDR6 platform should be assumed from development announcements. | Not established |
MT/s means million transfers per second. It is an interface transfer rate, not a direct measure of application performance.
A memory standard is one layer in a complete system. Headline bandwidth omits several constraints.
Channel count, bus width, protocol overhead, access patterns, controller scheduling, and workload locality determine how much theoretical bandwidth becomes useful.
A faster interface can move more data while first-access latency remains unchanged or increases. Final timings, controller behavior, and workload characteristics matter.
Production systems require a compatible controller, PHY, package or socket, board routing, firmware, thermal design, validation, and qualified modules.
These diagrams show logical data widths at a high level. They are not pinouts and omit ECC, command, address, and signaling details.
DDR5 divides the traditional module data width into two independently addressable 32-bit sub-channels, excluding ECC bits.
A 4 × 24-bit arrangement appears in industry reporting. Treat it as a direction under development, not a final universal module definition.
Buy for the platform and workload you need. DDR6 will require compatible processors, controllers, boards or packages, firmware, and validated memory—not a module swap into a DDR5 system.