CL Numbers, Nanoseconds, and What Actually Matters
The conversion takes one line of arithmetic. The conclusion most guides draw from it is wrong.
Quick Summary · TL;DR
True latency in nanoseconds = CL × 2000 ÷ data rate. By that measure DDR4-3600 CL16 (8.89ns) beats DDR5-6000 CL30 (10.00ns). It still loses in practice, because the DDR5 kit moves 67% more data per second and because latency is only part of what memory does for a game.
| Kit | CL | True latency | Peak bandwidth per stick |
|---|---|---|---|
| DDR4-3200 | CL16 | 10.00ns | 25.6 GB/s |
| DDR4-3600 | CL16 | 8.89ns | 28.8 GB/s |
| DDR4-3600 | CL18 | 10.00ns | 28.8 GB/s |
| DDR5-5600 | CL36 | 12.86ns | 44.8 GB/s |
| DDR5-6000 | CL30 | 10.00ns | 48.0 GB/s |
| DDR5-6000 | CL28 | 9.33ns | 48.0 GB/s |
| DDR5-6400 | CL32 | 10.00ns | 51.2 GB/s |
| DDR5-8000 | CL38 | 9.50ns | 64.0 GB/s |
The arithmetic
true latency (ns) = CL × 2000 ÷ data rate (MT/s)
The 2000 is not magic. A DDR bus transfers on both clock edges, so a 6000 MT/s kit runs a 3000MHz clock, one cycle is 1/3000 of a microsecond, and CAS latency is counted in clock cycles. Multiplying cycles by cycle time and converting to nanoseconds collapses to the line above. Our RAM latency calculator does nothing more than this.
Where the usual conclusion goes wrong
Plenty of guides stop at the table, notice that DDR4-3600 CL16 shows a lower number than DDR5-6000 CL30, and conclude that DDR4 is still the better gaming memory. That reasoning treats latency as the only variable, and it is not even the dominant one.
- Bandwidth roughly doubled. 28.8 GB/s against 48.0 GB/s per stick. Modern engines stream a great deal of geometry and texture data; the wider pipe matters more than a 1.1ns difference in first-word access.
- DDR5 has two independent 32-bit sub-channels per stick. A DDR4 stick has one 64-bit channel. More outstanding requests in flight hides latency that the ns figure cannot express.
- Platform memory controllers are tuned for it. On AM5 the widely-recommended DDR5-6000 target exists because it keeps the memory controller and fabric clocks in a 1:1 ratio; pushing further can raise the data rate while making effective latency worse.
- Capacity beats both. 32GB of DDR5-5600 CL36 will outperform 16GB of anything in a 2026 title that wants the headroom.
The honest summary: within one memory generation, lower nanoseconds is a real advantage and worth a modest premium. Across generations the comparison does not transfer, and using it to justify staying on an older platform is a mistake.
What to actually buy
| Platform | Sensible target | Why |
|---|---|---|
| AMD AM5 | DDR5-6000 CL30, 2 × 16GB | Keeps controller and fabric clocks synchronous; the widely validated sweet spot. |
| AMD AM4 | DDR4-3600 CL16, 2 × 16GB | The same 1:1 logic one generation earlier. |
| Intel 12th–14th Gen (DDR5) | DDR5-6400 CL32, 2 × 16GB | The controller tolerates higher data rates than AM5; extra speed is usable. |
| Any platform, tight budget | Capacity first, then speed, then CL | Running out of memory costs far more frames than a slower kit does. |
Two sticks, not four, wherever you have the choice. Four sticks load the memory controller harder and frequently mean settling for a lower validated data rate. And check your motherboard’s memory support list before buying — a kit’s advertised speed is an XMP or EXPO profile, not a guarantee on every board.
Convert your own kit
The RAM latency calculator converts any CL and data rate to nanoseconds, and the RAM buying guide has per-platform picks.
Sources & verification
Bandwidth figures are theoretical peaks (data rate × 8 bytes per stick). Advertised kit speeds require an XMP or EXPO profile and board support. Last reviewed September 12, 2026.