HBM (High Bandwidth Memory) chips come in various capacities depending on the specific generation and variant. Here’s a detailed overview:

1. HBM1 (First Generation):
- Capacity per chip: HBM1 chips typically have a capacity of 1 GB per stack.
- Stacks per device: Devices can have up to 4 stacks, resulting in a maximum capacity of 4 GB per device.
- Example: AMD%27s Radeon R9 Fury X GPU uses HBM1 with 4 GB of total memory (4 stacks x 1 GB each).

2. HBM2 (Second Generation):
- Capacity per chip: HBM2 chips can have capacities ranging from 2 GB to 8 GB per stack.
- Stacks per device: Devices can have up to 8 stacks, resulting in a maximum capacity of 16 GB to 64 GB per device, depending on the number and capacity of the stacks.
- Example: NVIDIA’s Tesla V100 and AMD’s Radeon VII GPUs both use HBM2 with up to 16 GB of total memory (4 stacks x 4 GB each for the Tesla V100 and 4 stacks x 8 GB each for the Radeon VII).

3. HBM2E (Second Generation Enhanced):
- Capacity per chip: HBM2E chips have higher capacities than HBM2, typically ranging from 8 GB to 16 GB per stack.
- Stacks per device: Devices can have up to 12 stacks, resulting in a maximum capacity of 96 GB per device (12 stacks x 8 GB each).
- Example: AMD’s Radeon RX 6900 XT GPU uses HBM2E with 16 GB of total memory (2 stacks x 8 GB each).

4. HBM3 (Third Generation):
- Capacity per chip: HBM3 is expected to further increase the capacity per stack, potentially reaching up to 24 GB per stack.
- Stacks per device: The total capacity per device will depend on the number and capacity of the stacks used in future products.

Additional Details:

- Stack Configuration: HBM chips are organized into multiple vertical stacks, typically ranging from 2 to 12 stacks per device.
- Bus Width: Each HBM2 and HBM2E stack has a wide bus width of 1024 bits, which contributes to its high bandwidth.
- Energy Efficiency: HBM chips are known for their energy efficiency, which is achieved through reduced power consumption per bit transferred compared to other memory types.

In summary, the capacity of HBM chips has increased significantly over different generations, from 1 GB per stack in HBM1 to potentially up to 16 GB or more per stack in HBM2E. This increase in capacity, combined with high bandwidth and low latency, makes HBM a preferred choice for high-performance applications such as GPUs in gaming, artificial intelligence, and high-performance computing.

icDirectory Limited | https://www.icdirectory.com/a/blog/what-is-the-capacity-of-hbm-chips.html