A Comprehensive Comparison of HBM

1. Introduction to HBM

HBM, or High Bandwidth Memory, is a revolutionary technology in the field of memory storage. It is characterized by its vertical stacking of multiple DRAM Die, connected through TSV (Through-Silicon Via) technology to the logic Die, enabling a small size with high bandwidth and high transmission speed. This technology has become a mainstream solution for the GPU memory of high-performance AI servers.

2. Evolution and Versions of HBM

HBM has gone through several iterations, from the initial HBM1 launched in 2014 by AMD and SK Hynix to the current HBM3 and its extended version HBM3e. HBM1 provided 128GB/s bandwidth and 4GB memory with a 4-layer die stack, significantly outperforming GDDR5 of the same period. The latest HBM3e offers an impressive transmission speed of up to 8Gbps and 16GB memory, first released by SK Hynix and expected to enter mass production in 2024.

3. Applications of HBM

HBM is mainly applied in AI servers. The latest generation HBM3e is equipped in NVIDIA's H200 released in 2023. With the growth in the shipment of AI servers, which is projected to exceed 2 million units by 2026 with a compound annual growth rate of 29% based on Trendforce data in 2022, the demand for HBM is experiencing an explosive growth. Coupled with the increase in the average HBM capacity of servers, it is estimated that the market size will reach approximately 15 billion US dollars in 2025, with a growth rate exceeding 50%.

4. Market Share and Suppliers of HBM

The supply of HBM is mainly concentrated among the three major storage original factories: SK Hynix, Samsung, and Micron. According to Trendforce data in 2023, SK Hynix is expected to have a market share of 53%, Samsung 38%, and Micron 9%.

5. Comparison with Other Memory Technologies

When comparing HBM with GDDR, HBM offers a revolutionary stacked memory architecture with TSV, providing a 1024-bit wide bus and superior bandwidth per watt, making it highly suitable for workloads such as AI and HPC. In contrast, GDDR6 relies on a more traditional planar architecture with 16-bit channels and high clock speeds, excelling in scenarios where speed and capacity are crucial, such as gaming and graphics.

6. Consistency Comparison and Verification of HBM and Related Technologies

In the consistency comparison of quantized.bc and hbm, the most complex part is the verification of stage3.bc and stage3.hbm. When using hb_verifier for comparison, certain information needs to be focused on. Input data preparation for this comparison can be done using tools such as hb_model_info and hrt_model_exec in specific environments.