SK Hynix Samples 24GB HBM3 Modules: Up to 819 GB/s
SK hynix is announced Announced that it developed the industry’s first 12-layer 24GB HBM3 memory stack, offering both high density and extreme bandwidth of 819 GB/s. The 12-Hi HBM3 product maintains the same height as the company’s existing HE 8-layer HBM3 product, making it easier to deploy.
SK Hynix’s 24GB HBM3 Known Good Stacked Die (KGSD) product places twelve 16Gb memory devices connected using Through Silicon Vias (TSV) on a base layer with a 1024-bit interface . The device features a data transfer rate of 6400 MT/s, offering 819.2 GB/s of bandwidth across the 24GB HBM3 module.
Depending on the actual memory subsystem, such modules can deliver 3.2 TB/s to 4.915 TB/s of bandwidth for 96 GB of memory with a 4096-bit interface or 144 GB of memory with a 6140-bit interface, respectively. can be enabled. Nvidia’s H100 NVL (the most advanced HBM3 implementation to date) features 96 GB of memory with 3.9 TB/s bandwidth on each of his two GH100 computing GPUs.
Stacking 12 layers of HBM DRAM is difficult for several reasons. First, it is difficult to drill 60,000+ TSV holes in a package and connect all 12 layers. Second, a 12-Hi HBM DRAM package cannot be physically higher than an 8-Hi HBM KGSD (usually 700-800 microns, 720 microns This makes mounting such an HBM3 KGSD next to a fixed height CPU or GPU very complicated (if possible). To that end, DRAM makers such as SK Hynix either reduce the thickness of individual DRAM layers without sacrificing yield or performance (which poses many challenges), or reduce the gap between layers to reduce the thickness of the base layer. should be reduced.
SK Hynix used Advanced Mass Reflow Molded Underfill (MR-MUF) encapsulation technology, including Mass Reflow (MR) chip attachment, to build the 12-Hi HBM3 product flush with the 8Hi HBM3 device, Reduce die-to-die spacing using base layer and mold underfill (MUF) processes.
SK Hynix provided samples of its 24GB HBM3 product to several eagerly-anticipated customers. We are currently evaluating the performance of the product and plan to start mass production in the second half of the year.