SK Hynix Ships 12-Layer HBM4E Samples for AI

What happened?

Last Thursday, SK Hynix announced it has begun shipping samples of its next-generation high-bandwidth memory (HBM), called HBM4E, to major clients in the artificial intelligence sector. This chip represents a significant generational leap: it uses a 12-layer stack reaching 48 GB capacity, operates at speeds up to 16 Gbps per pin, and offers improved energy efficiency over the previous HBM3E generation. According to The Next Web, SK Hynix has not disclosed the names of clients receiving the samples, but speculation includes NVIDIA and AMD, the leading AI accelerator manufacturers. The company indicated that mass production is planned for 2026, suggesting these initial shipments are for validation and future product design.

Why is this important?

HBM memory is a critical component in AI accelerators like NVIDIA and AMD GPUs, enabling fast access to large data volumes essential for training and running large language models (LLMs) and other AI workloads. The advance to 12 layers and 48 GB doubles the typical capacity of previous generations (HBM3E offered up to 24 GB with 8 layers), translating to the ability to handle larger, more complex models without bottlenecks. For example, models like GPT-4 or Gemini require tens of gigabytes of high-speed memory; with HBM4E, entire models could be loaded into a single GPU, reducing the need to split them across multiple chips and improving performance. Additionally, the improved energy efficiency (percentage not specified, but described as "improved") is key for data centers seeking to reduce operational costs and carbon footprint. In a context where AI energy consumption is a growing concern, any efficiency gain directly impacts economic and environmental viability.

Market consequences

This move reinforces SK Hynix's position as the HBM market leader, ahead of competitors like Samsung and Micron. According to market data, SK Hynix controls approximately 50% of the HBM market, followed by Samsung with 40% and Micron with 10%. Early shipment of HBM4E samples allows SK Hynix to get ahead of rivals, who are still developing their own 12-layer solutions. Samsung has announced plans to produce HBM4 in 2025 but has not yet shown an equivalent product to HBM4E. Micron, meanwhile, has focused on HBM3E and has not revealed plans for HBM4. Given that demand for high-bandwidth memory has surged with the rise of generative AI — the HBM market is expected to grow from $4 billion in 2023 to over $20 billion in 2028 — having early samples allows clients to integrate the technology into their upcoming products, potentially creating competitive advantages for SK Hynix. However, mass production will not begin until 2026, giving competitors time to catch up. Additionally, SK Hynix's manufacturing capacity will be a limiting factor; the company has invested in new production lines in South Korea, but demand could outstrip supply.

What readers should know

For companies and developers, this means upcoming AI chips will be even more powerful and efficient. With 48 GB per stack, GPUs could double their current HBM memory (e.g., the NVIDIA H100 has 80 GB with five 16 GB HBM3 stacks; with HBM4E, configurations up to 96 GB with fewer stacks could be achieved, freeing up substrate space). However, actual availability will depend on SK Hynix's manufacturing capacity and adoption by hardware manufacturers. The names of clients receiving samples have not been disclosed, but speculation includes NVIDIA and AMD. It is also possible that custom chip design companies like Google (TPU) or Amazon (Trainium) are evaluating these samples. A key aspect is that HBM4E uses an improved interface that may require changes in GPU memory controllers, implying integration effort. Additionally, improved energy efficiency could enable less aggressive cooling systems, reducing data center costs. Compared to previous events, like the jump from HBM2 to HBM3, this advance is more incremental in speed (from 6.4 Gbps to 8 Gbps in HBM3, and now to 16 Gbps) but significant in capacity (from 8 layers to 12). This reflects the industry trend of prioritizing capacity over pure speed, as AI models grow ever larger.

"SK Hynix continues to push the boundaries of HBM memory, and the 12-layer HBM4E samples represent a major milestone for the AI industry."

In summary, SK Hynix's announcement is a step forward in the race for high-performance memory for AI, with direct implications for future accelerator performance, competition among memory manufacturers, and data center evolution. The coming months will be crucial to see how Samsung and Micron respond, and whether SK Hynix's clients can integrate this technology into their roadmaps. For AI developers, this means memory constraints could ease, enabling larger and more complex models without the need for multi-GPU infrastructure.