SK hynix Begins Sampling HBM3e, Volume Production Planned For H1 2024

by Anton Shilov on August 21, 2023 4:15 PM EST

• Posted in
• Memory
• GPUs
• DRAM
• SK Hynix
• HBM3
• HBM3E

SK hynix connected Monday announced that it had completed first improvement of its first HBM3E representation stacks, and has begun sampling nan representation to a customer. The updated ("extended") type of nan precocious bandwidth representation exertion is scheduled to statesman shipping successful measurement successful nan first half of adjacent year, pinch hardware vendors specified arsenic NVIDIA already lining up to incorporated nan representation into their HPC-grade compute products.

First revealed by SK hynix backmost astatine nan extremity of May, HBM3E is an updated type of HBM3 that is designed to timepiece higher than existent HBM3, though circumstantial clockspeed targets look to alteration by manufacturer. For SK hynix, arsenic portion of today's disclosure nan institution revealed that their HBM3E representation modules will beryllium capable to deed information transportation rates arsenic precocious arsenic 9 GT/sec, which translates to a highest bandwidth of 1.15 TB/sec for a azygous representation stack.

Curiously, SK hynix has yet to uncover thing astir nan planned capacity for their next-gen memory. Previous investigation from TrendForce projected that SK hynix would wide nutrient 24 GB HBM3E modules successful Q1 2024 (in clip to reside applications for illustration NVIDIA's GH200 pinch 144GB of HBM3E memory), boosting capacity complete today's 16GB HBM3 stacks. And while this still seems apt (especially pinch nan NV announcement), for now it remains unconfirmed.

TrendForce HBM Market Projections (Source: TrendForce)

Meanwhile, nan SK hynix besides confirms that its HBM3E stacks are group to usage its Advanced Mass Reflow Molded Underfill (MR-RUF) technology to trim their power dissipation by 10%. But thermals is not nan only use MR-RUF tin provide. MR-RUF implies nan usage of an improved underfill betwixt layers, which improves thermals and reduces thickness of HBM stacks, which allows nan building of 12-Hi HBM stacks that are only arsenic gangly arsenic 8-Hi modules. This does not automatically mean that we are dealing pinch 12-Hi HBM3E stacks here, of course.

At present, SK hynix is nan only precocious measurement shaper of HBM3 memory, giving nan institution a very lucrative position, particularly pinch nan detonation successful request for NVIDIA's H100 and different accelerators for generative AI. And while nan improvement of HBM3E is meant to thief SK hynix support that lead, they will not beryllium nan only representation vendor offering faster HBM adjacent year. Micron besides threw their chapeau into nan ringing past month, and wherever those 2 companies are, Samsung is ne'er excessively acold behind. In fact, each 3 companies look to beryllium outpacing JEDEC, nan statement that is responsible for standardizing DRAM technologies and various representation interfaces, arsenic that group has still not published finalized specifications for nan caller memory.