Field notes
If you learn one hardware lesson from this whole site, make it this one: AI performance is usually decided by memory, not compute. Three companies make the memory that matters, it's been sold out for years, and every GPU roadmap on this site bends around it.
Start with these three
HB
High-Bandwidth Memory (HBM)
Read this first. Stacked memory is why modern AI chips work at all.
HB
HBM4
The 2026 generation — wider, faster, and already spoken for.
SK hynix
The quiet winner of the entire AI boom.
SK hynix
The quiet winner of the entire AI boom.
Memory technologies
9
HB
High-Bandwidth Memory (HBM)
Stacked DRAM wired straight to the GPU — the scarcest commodity in AI.
HB
HBM3E
The 2024–26 workhorse: up to 36 GB per stack in 12-high form.
HB
HBM4
2026's generation: a 2,048-bit interface and logic-die base from the foundries.
CX
CXL
Compute Express Link: pooled and expanded memory over PCIe wiring.
D&
DDR5 & LPDDR5X
The system memory feeding CPUs — and, via Grace, the GPUs too.
GD
GDDR7
Graphics memory conscripted for cost-efficient inference.
HB
HBM2
The 2016–2020 generation that powered V100 and A100-40GB.
HB
HBM2E
The refined HBM2 behind A100-80GB, Gaudi and Maia.
HB
HBM3
The H100/MI300X generation: ~0.8 TB/s per stack.