SK Hynix stated that it had began quantity manufacturing of its HBM3E reminiscence and would provide it to a buyer in late March. The South Korean firm is the second DRAM producer to announce mass manufacturing of HBM3E, so the market of ultra-high-performance reminiscence may have some competitors, which is nice for corporations that plan to make use of HBM3E.
According to specs, SK Hynix’s HBM3E identified good stack dies (KGSDs) characteristic information switch charges as much as 9.2 GT/s, a 1024-bit interface, and a bandwidth of 1.18 TB/s, which is massively greater than the 6.four GT/s and 819 GB/s supplied by HBM3. The firm doesn’t say whether or not it mass produces 8Hi 24GB HBM3E reminiscence modules or 12Hi 36GB HBM3E gadgets, however it should probably start its HBM3E ramp from lower-capacity merchandise as they’re simpler to make.
We already know that SK Hynix’s HBM3E stacks make use of the corporate’s superior Mass Reflow Molded Underfill (MR-RUF) know-how, which guarantees to scale back warmth dissipation by 10%. This know-how includes using an enhanced underfill between DRAM layers, which not solely improves warmth dissipation but additionally reduces the thickness of HBM stacks. As a outcome, 12-Hi HBM stacks could be constructed which might be the identical top as 8-Hi modules. However, this doesn’t essentially indicate that the stacks at present in mass manufacturing are 12-Hi HBM3E stacks.
Although the reminiscence maker doesn’t formally verify this, SK Hynix’s 24GB HBM3E stacks will arrive simply in time to handle NVIDIA’s Blackwell accelerator household for synthetic intelligence and high-performance computing functions.
“With the success story of the HBM enterprise and the robust partnership with clients that it has constructed for years, SK Hynix will cement its place as the full AI reminiscence supplier,” stated Sungsoo Ryu, Head of HBM enterprise at SK Hynix. As a outcome, NVIDIA may have entry to HBM3E reminiscence from a number of suppliers with each Micron and SK Hynix.
Meanwhile, AMD just lately confirmed that it was trying ahead to increasing its Instinct MI300-series lineup for AI and HPC functions with higher-performance reminiscence configurations, so SK Hynix’s HBM3E reminiscence may be used for this.
