Memory systems

HBM Catalog

Version-level comparison of high bandwidth memory generations used around GPUs, AI accelerators, and HPC packages.

How to read this table

HBM numbers are usually quoted per stack. Accelerator-level bandwidth depends on the number of stacks and the vendor's memory controller implementation.

HBM3E is a product evolution around the HBM3 family, so published limits vary by supplier and product generation.

Metric HBMHBM2HBM2EHBM3HBM3EHBM4HBM4E
Bus Width (Interface) 1024-bit1024-bit1024-bit1024-bit1024-bit2048-bit2048-bit
Pin Speed (Signaling) 1 Gb/s per pin2 Gb/s per pin3.6 Gb/s per pin6.4 Gb/s per pin9.6 Gb/s per pin8 Gb/s per pin16 Gb/s per pin
Peak Bandwidth per Stack 128 GB/s per stack256 GB/s per stack461 GB/s per stack819 GB/s per stack1.3 TB/s per stack2 TB/s to 3.3 TB/s4 TB/s
Capacity per stack 16 GB16 GB36 GB64 GB64 GB64 GB64GB
Channels na8 independent channels8 independent channels16 independent channels16 independent channels32 independent channels32 independent channels
Stack heights 8-Hi8-Hi12-Hi16-Hi16-Hi16-Hi16-Hi
Used by Prior generation AI accelerators and HPC GPUsPrior generation AI accelerators and HPC GPUsPrior generation AI accelerators and HPC GPUsH100-class and MI300-class AI acceleratorsH200, B200, MI325X, MI355XNext-generation AI and HPC acceleratorsNext-generation AI and HPC accelerators
Sources