SK Hynix said that its HBM3E memory has begun mass production and will be available to customers by the end of March. This Korean company will be the second DRAM manufacturer to announce mass production of HBM3E, so there will be some competition in the ultra-high-performance memory market, which is good for companies planning to use HBM3E.
According to the specifications, SK Hynix’s HBM3E Known Good Stacked Die (KGSD) has a data transfer rate of 9.2 GT/s, a 1024-bit interface, and a bandwidth of 1.18 TB/s, which is much higher than the 6.4 GT/s and HBM3 offers 819 GB/s . The company hasn’t revealed whether it will mass-produce 8Hi 24GB HBM3E memory modules or 12Hi 36GB HBM3E devices, but it will likely start with lower-capacity products as those are easier to manufacture.
We already know that SK Hynix’s HBM3E stack uses the company’s advanced large-scale reflow molded underfill (MR-RUF) technology, which promises to reduce heat dissipation by 10%. The technology involves using enhanced underfill between DRAM layers, which not only improves heat dissipation but also reduces the thickness of the HBM stack. Therefore, it is possible to build a 12-Hi HBM stack with the same height as an 8-Hi module. However, this does not necessarily mean that the stack currently in production is the 12-Hi HBM3E stack.
Although the memory maker has not officially confirmed this, SK Hynix’s 24GB HBM3E stack will be available in time to handle the demand for Nvidia’s H200 GPUs for artificial intelligence and high-performance computing applications.
“Building on the success story of the HBM business and the strong partnerships it has built with customers over the years, SK hynix will solidify its position as an overall AI memory provider,Sungsoo Ryu, head of HBM business of SK hynix said. Therefore, Nvidia will be able to use HBM3E memory from Micron and SK Hynix.
Meanwhile, AMD recently confirmed that it is looking to expand its Instinct MI300 series product line for AI and HPC applications with higher-performance memory configurations, so SK Hynix’s HBM3E memory could also be used for this purpose.
Source: SK hynix