The one number that compares every kit
Kit labels give timings in clock cycles, and a cycle isn't the same length on every kit. True latency fixes that by converting the wait to nanoseconds: CL × 2000 ÷ the data rate in MT/s. Run popular kits through it and the CL numbers stop telling the story:
- DDR3-1600 CL9: 11.3 ns
- DDR4-3200 CL16: 10.0 ns
- DDR4-3600 CL16: 8.9 ns
- DDR5-4800 CL40: 16.7 ns
- DDR5-5600 CL36: 12.9 ns
- DDR5-6000 CL30: 10.0 ns
- DDR5-6400 CL32: 10.0 ns
- DDR5-8000 CL38: 9.5 ns
Reading the gaps
A tuned DDR4 kit and a tuned DDR5 kit sit close together. DDR4-3600 CL16 waits 8.9 ns and DDR5-6000 CL30 waits 10.0 ns, a gap of 1.1 ns. The outlier is DDR5-4800 CL40. It waits 67% more than DDR4-3200 CL16, which is why early DDR5 didn't win on latency.
Buying speed doesn't always buy latency either. DDR5-8000 CL38 has a data rate 33% above DDR5-6000 CL30, yet its wait of 9.5 ns is only 5% below the 10.0 ns of the slower kit. The extra speed goes into bandwidth.
What true latency leaves out
It's the DRAM's wait, not the whole trip. A CPU request passes through its caches and the memory controller before it reaches the module, so the Memory and Storage Throughput Visualizer lists a full RAM access at 50 to 80 ns. Next to that, the 1.1 ns between a tuned DDR4 kit and a tuned DDR5 kit is a small slice.
The number also assumes the right row is already open. Opening a row costs tRCD and closing the wrong one costs tRP, and a random access pays both. And bandwidth is a separate score: 48.0 GB/s per channel at DDR5-6000 against 28.8 GB/s for DDR4-3600.
Which kit to shop on
Pick the lowest true latency among the kits your CPU and board can run, then weigh bandwidth, price and capacity. Make sure the kit's profile is switched on first. A kit can boot at a JEDEC setting until you enable XMP or EXPO in the BIOS, and then the figures on its label don't apply.