SK hynix shows HBM4, HBM3E and liquid SSDs
SK hynix demoed next‑gen HBM4 and HBM3E memory and unveiled liquid‑cooled SSDs developed with NVIDIA — moves that tighten the memory+storage stack for large AI training and inference farms. These components are being positioned as critical to squeezing latency and power out of AI datacenters. (sg.finance.yahoo.com)
SK hynix says its HBM4 design uses a 2,048‑bit I/O and pushes per‑pin speeds above 10 Gb/s—exceeding JEDEC’s 8 Gb/s baseline—while the company claims over 40% improved power efficiency and up to a 69% uplift in AI service performance versus prior generations. (videocardz.com) The company plans to manufacture HBM4 using an Advanced MR‑MUF (mass reflow molded underfill) package and a 1bnm (fifth‑generation 10 nm‑class) DRAM process, according to SK hynix’s development lead Joohwan Cho. (videocardz.com) SK hynix’s 12‑layer HBM3E stacks entered volume production in late 2024 at up to 36 GB per stack and 9.6 Gb/s transfer rates, yielding roughly 1.22 TB/s peak bandwidth per module and about 9.83 TB/s for an eight‑module memory subsystem. (tweaktown.com) A joint SK hynix–NVIDIA SSD program carries internal names “Storage Next” (NVIDIA) and “AI‑NP” (SK hynix) and targets an architectural leap to roughly 100 million IOPS—about ten times current top SSD random IOPS—with a prototype targeted by the end of 2026 and broader goals toward 2027. (techspot.com) CRN reports the same SSD effort is being pitched to support NVIDIA’s Rubin CPX inference GPUs (Rubin CPX ships with 128 GB of GDDR7 and is expected in the second half of 2026), tying the storage proof‑of‑concept schedule to upcoming inference hardware. (crn.com) SK hynix showed its HBM3E and HBM4 product models at GTC 2026 and has indicated multiple accelerator partners—NVIDIA, AMD and Intel—are planning HBM4 support in forthcoming server GPUs and accelerators. (news.skhynix.com)
