HBM3E

HBM3E is the extended version of HBM3, the high-bandwidth memory (HBM) that AI accelerators carry next to the processor; SK hynix calls it the fifth generation of HBM, after HBM, HBM2, HBM2E and HBM3.¹ Like other HBM it stacks DRAM dies and links them with through-silicon vias, and Micron says HBM needs more wafers and more cleanroom space than conventional DRAM to make the same number of bits on the same technology node.² SK hynix said on March 19, 2024 that it had begun volume production of HBM3E, the first maker to do so by its own account, for supply to a customer from late March, and that the product processes up to 1.18 TB of data per second.¹

It is the memory on most of the accelerators in AI Accelerators With the Most HBM per Chip (October 2026): ten of the thirteen ranked parts use HBM3E. They run from AMD’s Instinct MI350X and MI355X, tied for #1 at 288 GB and 8 TB/s,³ through NVIDIA’s Blackwell GPUs at 186 GB in the GB200 NVL72 and 180 GB in the HGX B200,⁴ to NVIDIA’s H200 at 141 GB and 4.8 TB/s.⁵ The third part tied at #1, NVIDIA’s Rubin, is the one ranked accelerator on HBM4 instead.

Counterpoint Research estimated in September 2026 that most HBM revenue in the second quarter of 2026 came from HBM3E, with HBM4 shipments expected to become visible in earnest only from the second half of the year.⁶ For HBM in general, TrendForce reports that GPU and ASIC makers are weighing 8-high stacks, which use fewer DRAM dies than 12-high ones, and projects 8-high parts to carry a premium of roughly 10–20% per gigabit over 12-high in 2027 because each stack still needs a base die.⁷ Both figures are in The Memory Crunch, by the Numbers.