DDR5-6000 Dual Channel

DDR5-6000 in a dual-channel configuration has a theoretical maximum of 96 GB/s (6000 MT/s × 8 bytes × 2 channels = 96,000 MB/s). Real sustained bandwidth is lower and depends on the CPU, the memory controller and the settings. One indirect data point: a public llama.cpp result on a Ryzen 9 7950X implies about 60 GB/s of effective bandwidth during token generation (see the Ryzen 9 7950X entry). That is arithmetic on one benchmark whose memory speed is not stated, not a measurement of DDR5-6000.

Why this matters for hybrid offload

When running llama.cpp with partial GPU offload (--n-gpu-layers less than the full layer count), the layers left on the CPU are bound by system memory bandwidth, not GPU VRAM bandwidth. A 70B Q4_K_M model (a 39.6 GB file) with roughly half its layers on a 24GB GPU still bottlenecks on system RAM for the CPU-resident half. Overall tokens/sec is limited by the slower half, and system RAM is almost always slower than GPU VRAM, which runs from about 288 GB/s on a mid-range card to about 1,008 GB/s on an RTX 4090.

Single vs. dual channel: verify this, don’t assume it

A common, easy-to-miss misconfiguration: motherboards and RAM kits don’t always default to dual-channel operation, depending on DIMM slot population. Two sticks in the wrong slot pairing run as single-channel, which has half the theoretical bandwidth (about 48 GB/s), silently halving CPU-side inference throughput with no error message anywhere.

# Linux: list the populated DIMM slots and their sizes
sudo dmidecode --type memory | grep -A2 "Locator:"

This shows which slots are populated, not the channel mode itself. Compare it with your motherboard manual’s DIMM pairing guidance. On Windows, the Memory tab of a tool such as CPU-Z typically shows the channel mode directly.

Upgrade path

DDR5-8000 dual-channel kits have a theoretical bandwidth of 128 GB/s (8000 × 16 bytes), but CPU memory controller support and motherboard QVL (qualified vendor list) compatibility become real constraints at those speeds. Verify CPU and motherboard support before purchasing memory faster than what is officially validated for the platform.

Known limitation

The 96 GB/s figure is a calculation of the theoretical peak, not a measured result, which is why this entry keeps the “estimated” label. See /methodology for what that means on this site.