SK Hynix announces first 16-layer HBM3E memory, 48GB

SK Hynix this week announced another advancement in its HBM3E memory technology. Now its products reach up to 16 layers, with 3GB per layer, for a total of 48GB per stack. This means that the manufacturer can now help create AI accelerators with up to 384GB of capacity, in 8-stack configurations.

The announcement marks a step forward in the race of the three memory giants, as in September the company already began mass production of HBM3E in 12 layers. SK Hynix will begin offering test samples of its 16-layer memory early next year.

With the storage density offered by the new generation HBM3E, SK Hynix promises up to 18% improvement in machine training. Furthermore, the gain in performance for inference can reach up to 32%, according to the manufacturer.

Despite the advances promised by the company, one has to wonder how long the search for the new version of HBM3E will last. For example, the next generation of NVIDIA GPUs, currently known as Rubin, are expected to use HBM4 in their manufacturing. SK Hynix itself even emphasized that its development of the new technology continues at full speed.

SK Hynix memory solutions go beyond HBM3E

The South Korean memory company spoke about the new version of its technology for HBM3E at its own event, the SK AI Summit 2024. During the presentation, the company also offered updates on its advances in other solutions for AI servers.

SK Hynix highlighted, for example, that it is working on high-capacity QLC PCIe 6.0 SSDs for servers and data centers. In addition, the company is also advancing in the development of UFS 5.0 storage for mobile devices.

More about HBM memories:

Finally, the company also commented on its advances in PNM and PIM technologies – Processing Near Memory and Processing In Memory, respectively. As the names indicate, they are solutions for processing part of the data directly in memory, without the delay of having to communicate with the processor.

Source: Tom’s Hardware