Memory technology: DRAM, SRAM and HBM
In short: Computers use different kinds of memory that differ in speed, price and density. The closer to the processor, the faster, smaller and more expensive.
The memory pyramid
- Registers in the processor: a few bytes, available instantly.
- SRAM (cache): fast, needs several transistors per bit, so expensive and small (kilo- to megabytes), see Caches.
- DRAM (main memory): one transistor and a tiny capacitor per bit. Dense and cheap (gigabytes), but the capacitor loses charge and must be refreshed every few milliseconds. Hence “dynamic”.
- Flash/SSD: keeps data without power, slower (NAND-Flash, SSDs).
- Hard disk: even larger, even slower.
DDR memory
DDR (double data rate) transfers data on rising and falling clock edges. Generations: DDR4, DDR5 (more bandwidth, two sub-channels per module, error correction in the chip). The number in “DDR5-6000” gives millions of transfers per second. Important values: bandwidth (how much per second) and latency (how long an access takes). Several channels (dual channel) double the bandwidth.
Server modules have ECC (error correction), which detects and fixes random bit errors.
Special memory
- GDDR: for graphics cards, high bandwidth.
- HBM (high bandwidth memory): memory chips are stacked and connected through very many lines directly next to the processor. Extremely high bandwidth at lower energy, but expensive. Standard in AI accelerators because language models mostly wait for memory, see AI chips compared.
- LPDDR: power-saving for phones and laptops.
Good to know
- More RAM only helps if the existing one is short (memory management and the OOM killer).
- Two identical modules are better than one for both channels.
- Main memory is volatile: when switched off the data is gone.
See also: RAM