GPUs and AI Development

Graphics processing units (GPUs) have become widely recognized due to their growing role in AI development. However, a lesser-known but critical technology is also gaining attention: high-bandwidth memory (HBM).

HBM is a high-density memory designed to overcome bottlenecks and maximize data transfer speeds between storage and processors. AI chipmakers like Nvidia rely on HBM for its superior bandwidth and energy efficiency. Its placement next to the GPU’s processor chip gives it a performance edge over traditional server RAM, which resides between storage and the processing unit. HBM’s ability to consume less power makes it ideal for AI model training, which demands significant energy resources.

However, as the AI landscape transitions from model training to AI inferencing, HBM’s widespread adoption may slow. According to Gartner’s 2023 forecast, the use of accelerator chips incorporating HBM for AI model training is expected to decline from 65% in 2022 to 30% by 2027, as inferencing becomes more cost-effective with traditional technologies.

How HBM Differs from Other Memory HBM shares similarities with other memory technologies, such as graphics double data rate (GDDR), in delivering high bandwidth for graphics-intensive tasks. But HBM stands out due to its unique positioning. Unlike GDDR, which sits on the printed circuit board of the GPU, HBM is placed directly beside the processor, enhancing speed by reducing signal delays caused by longer interconnections. This proximity, combined with its stacked DRAM architecture, boosts performance compared to GDDR’s side-by-side chip design.

However, this stacked approach adds complexity. HBM relies on through-silicon via (TSV), a process that connects DRAM chips using electrical wires drilled through them, requiring larger die sizes and increasing production costs. According to analysts, this makes HBM more expensive and less efficient to manufacture than server DRAM, leading to higher yield losses during production.

AI’s Demand for HBM Despite its manufacturing challenges, demand for HBM is surging due to its importance in AI model training. Major suppliers like SK Hynix, Samsung, and Micron have expanded production to meet this demand, with Micron reporting that its HBM is sold out through 2025. In fact, TrendForce predicts that HBM will contribute to record revenues for the memory industry in 2025.

The high demand for GPUs, especially from Nvidia, drives the need for HBM as AI companies focus on accelerating model training. Hyperscalers, looking to monetize AI, are investing heavily in HBM to speed up the process.

HBM’s Future in AI While HBM has proven essential for AI training, its future may be uncertain as the focus shifts to AI inferencing, which requires less intensive memory resources. As inferencing becomes more prevalent, companies may opt for more affordable and widely available memory solutions.

Experts also see HBM following the same trajectory as other memory technologies, with continuous efforts to increase bandwidth and density. The next generation, HBM3E, is already in production, with HBM4 planned for release in 2026, promising even higher speeds.

Ultimately, the adoption of HBM will depend on market demand, especially from hyperscalers. If AI continues to push the limits of GPU performance, HBM could remain a critical component. However, if businesses prioritize cost efficiency over peak performance, HBM’s growth may level off.