Last updated:
HBM2 is the second generation of High Bandwidth Memory: stacked DRAM connected to a processor through a very wide interface. It is associated with GPUs and compute accelerators that need to move large amounts of data. Its absence from many consumer gaming cards does not mean the HBM approach failed.
How the stacked memory works
Rather than placing many separate memory packages around the processor on a circuit board, an HBM design places compact stacks close to the processor within an advanced package. Through-silicon vias (TSVs) provide vertical connections through stacked dies. An interposer or another suitable packaging structure links the stacks to the processor.
The combination of many data connections and short physical paths allows high aggregate bandwidth without relying solely on a high signaling rate on each connection. The exact bandwidth and capacity depend on the number of stacks, their specifications and the processor’s interface.
A concrete HBM2 example
NVIDIA’s Tesla P100 SXM2 datasheet lists 16 GB of HBM2 and 732 GB/s of memory bandwidth. This is a specification for that product, not a bandwidth promise for every HBM2 implementation. Other P100 variants have different capacity or bandwidth specifications.
NVIDIA’s Pascal architecture explanation shows how HBM2 and the GPU package were designed together. A consumer cannot replace an existing card’s GDDR chips with an HBM2 stack as an upgrade.
HBM2 versus GDDR memory
| Design choice | HBM2 | GDDR approach |
|---|---|---|
| Placement | Stacked dies close to the processor in an advanced package | Separate memory packages on the board |
| Bandwidth strategy | Very wide interface with many connections | Higher per-pin data rates across the card’s memory bus |
| Design considerations | Package integration, assembly, capacity and cost | Board routing, component placement, bus width and signaling |
| Typical use | Bandwidth-intensive compute and some historical graphics products | Many consumer graphics cards and other graphics products |
There is no universal latency or power multiplier that makes every HBM2 product better than every GDDR card. Compare measured workload performance, capacity and complete device power. Our GDDR6X versus GDDR6 guide explains the signaling distinction on the GDDR side.
Why manufacturers choose different memory
Package cost, manufacturing capacity, memory capacity, power and the intended workload all influence the choice. A gaming card also needs an appropriate GPU, cooling system and price; memory bandwidth alone does not determine frame rate. Compute workloads can have a different balance of bandwidth, capacity and power requirements.
HBM2 is not the same as HMB
Newer HBM generations have their own specifications and should not be described as HBM2. HMB, or Host Memory Buffer in an NVMe SSD, is another concept entirely: it lets a drive use allocated host memory for certain working data. Similar initials do not make the technologies interchangeable.