The short version
RAM is your CPU's workbench. Programs and game data are copied from your SSD into RAM because the CPU can reach it hundreds of times faster. It's fast, but it forgets everything when the power goes off.
- CapacityHow much fits on the workbench at once. Run out, and the PC falls back to the much slower SSD.
- SpeedHow fast data streams once it starts moving (the 6000 in DDR5-6000).
- LatencyHow long the CPU waits before the data starts moving (the CL number).
Where RAM sits: the speed ladder
Typical times for one small piece of data to reach the CPUYour CPU keeps the data it's using right now in tiny, very fast caches built into the chip. When the data isn't there, it has to go out to RAM, and if it isn't in RAM either, to the SSD. Press the button to race them.
| Where the data is | Typical wait | If 1 nanosecond were 1 second |
|---|---|---|
| L1 cache | About 1 ns | 1 second |
| L2 cache | About 3 ns | 3 seconds |
| L3 cache | About 10 ns | 10 seconds |
| RAM | About 70–80 ns | A little over a minute |
| NVMe SSD | About 50–100 microseconds | 14 to 28 hours |
| Hard drive | About 5–10 milliseconds | 2 to 4 months |
Inside a memory stick
- The stickA 16GB DDR5 stick usually carries 8 memory chips of 16 gigabits each.
- The chipSplit into banks that work independently: 32 on DDR5, 16 on DDR4.
- The bankA grid of about 65,000 rows, each 8,192 bits wide, plus a row buffer to read them into.
- The cellOne transistor and one tiny capacitor. Charged = 1, empty = 0. A stick has billions of them.
What happens when the CPU asks for data
Real timings, slowed down so you can watchEvery read goes through the same few steps, and each step has a fixed wait measured in clock cycles. These waits are the timings printed on every memory kit, like the 30-38-38 in "DDR5-6000 CL30-38-38". Try different kits and starting points.
- CL (CAS latency) is the first and most advertised timing: the wait between asking for a column and the data starting to come out.
- tRCD is the wait to open a row, and tRP the wait to close one. A fourth number on some boxes, tRAS, is the minimum time a row must stay open.
- Cycles aren't nanoseconds. Faster memory has shorter cycles, so DDR5-6000 at CL30 waits exactly as long (10 ns) as DDR4-3200 at CL16. Compare kits in nanoseconds: CL × 2000 ÷ speed.
- Row hits are the cheap case. The memory controller in your CPU reorders requests to reuse open rows, and having many banks lets other rows open while one is busy.
Why it's called "double data rate"
Memory runs on a clock signal that ticks up and down. Older memory moved data once per tick; DDR moves data on the way up and on the way down.
Channels: why two sticks beat one
Desktop CPUs talk to memory over two separate channels, A and B. Each one can move data at the same time as the other, but only if it has a stick in it. On DDR5, each stick is further split into two independent 32-bit sub-channels, so the controller can juggle more requests at once.
With two sticks, use the slots your motherboard manual recommends, usually A2 and B2 (the second and fourth slots from the CPU).
Why RAM forgets: refresh
Each cell's capacitor is so tiny that its charge leaks away in a fraction of a second. To keep your data, the memory constantly reads every row and writes it back. That's the "dynamic" in DRAM, and it's why RAM loses everything when the power goes off.
What this means when you buy
- Capacity first. Running out of RAM sends the PC to the SSD, which is around a thousand times slower. 32GB is the comfortable amount for gaming today; 16GB still works for lighter use.
- Two sticks, not one. A 2 × 16GB kit gives you both channels and double the bandwidth of a single 32GB stick.
- Look at latency in nanoseconds. A faster kit with a higher CL can be just as quick to respond. DDR5-6000 CL30 is the popular sweet spot for current AMD Ryzen CPUs.
- Turn on EXPO or XMP in the BIOS. Without it, most kits run at a slower standard speed, like the DDR5-4800 setting above.
Today: 32GB of DDR5 from …, and a DDR5-6000 CL30 32GB kit from ….