DDR4 RAM price history

ECC (error-correcting code) memory stores 8 extra check bits for every 64 bits of data, so a single bit that occasionally flips in memory is corrected automatically. That is why an ECC stick has one or two more chips than a regular stick of the same spec. Server memory (RDIMM) always has ECC. On a desktop, error correction only works if both the CPU and the motherboard support it. On a platform that doesn't, an ECC UDIMM usually still works, just without error correction. You don't need to buy ECC memory specifically to run local LLMs.

Desktops use unbuffered memory (UDIMM); servers use registered memory (RDIMM). An RDIMM has an extra register chip that takes load off the memory controller, so each channel can hold more sticks and each stick can be larger (32GB and 64GB are common). All RDIMMs have ECC. The trade-off is lower speeds and looser timings: server sticks are commonly DDR4-2400 or 2666, desktop sticks 3200 or 3600. The two types are not electrically compatible: an RDIMM won't work on a desktop motherboard, and a UDIMM won't work on a server motherboard that only takes RDIMMs.

The number before R is the rank count. A rank is a group of chips on the stick that are read and written together to make up 64 bits of data: 1R is single-rank, 2R is dual-rank. The number after x is the data width of each chip in bits. It takes 8 x8 chips or 16 x4 chips to make 64 bits, so x4 sticks can reach higher capacities, and most high-capacity server sticks are 2Rx4. The more ranks in a channel, the more likely the motherboard is to lower the memory speed when all slots are filled.

CL (CAS latency) is the number of clock cycles between a read command and the data coming out. Lower is faster. To get the real latency, convert it using the speed: nanoseconds = CL × 2000 ÷ speed in MT/s. DDR4-3200 CL16, common on desktops, works out to 10 ns; DDR4-2666 CL19, common on servers, to about 14 ns. When an LLM runs, every generated token means reading several to tens of GB of weights from memory. Bandwidth is what counts, and timings have very little effect.

Sticks with different speeds will work together, but they all run at the speed of the slowest stick, with the loosest timings. With different capacities, only the capacity that matches on both channels is interleaved in dual-channel mode, so bandwidth drops. Different brands usually work, but the safest choice is the same model from the same batch. Three combinations can't be mixed: desktop sticks (UDIMM) with server sticks (RDIMM), RDIMM with LRDIMM, and DDR4 with DDR5 (the notch is in a different place, so they don't fit). On a server, every channel needs identical sticks to reach full bandwidth.

A stick transfers 64 bits, or 8 bytes, at a time, so bandwidth per stick = speed × 8. DDR4-3200 gives 25.6 GB/s, which is what PC4-25600 on the label refers to. Platform bandwidth = bandwidth per stick × number of channels. A desktop has only 2 channels, so 4 sticks still give 51.2 GB/s: extra sticks add capacity, not bandwidth. EPYC 7002 / 7003 has 8 channels, which reach 204.8 GB/s when all are filled.

The number is the speed, and the letter after it is the JEDEC timing grade: 2133P is CL15, 2400T is CL17, 2666V is CL19, 2933Y is CL21 and 3200AA is CL22. PC4-21300 is the same speed written as bandwidth (2666 × 8 ≈ 21300 MB/s). In PC4-2666V-RB2, a string often seen in Xianyu listing titles, the R after the hyphen means RDIMM; U would be a desktop stick, S a laptop stick and L an LRDIMM. The letters and digits after that are the PCB revision, which you can ignore when buying.