HBM3 for the H100
On 8 June 2022 SK hynix began mass production of HBM3 and named the buyer for the first time: NVIDIA. 819 gigabytes a second per stack at 6.4 gigabits a second per pin, with supply matched to the shipping schedule of H100 systems in the third quarter of that year.
Why it matters
The atlas holds the H100 as an architecture with a unit built for one model shape. Here is the other half of the same machine, and the fact that it arrives from another country and another company: the accelerator stops being one supplier's product before anyone restricts it.
The figures come from two SK hynix releases. Per-pin speed rose from 1 gigabit a second in the first HBM generation to 6.4 in the fourth, giving 819 gigabytes a second per stack, which the company renders as 163 Full-HD movies of 5 gigabytes each, every second. Capacity: 24 gigabytes, the largest in the industry at the time, and a 16-gigabyte part; HBM2E topped out at 16. On-die error correction was added. Seven months separate the October 2021 development announcement from mass production. What this record does not claim. "More than 90 per cent of the initial supply for Hopper systems" is in neither release, and it does not enter. Nor does MR-MUF, the moulding compound the report used to explain heat removal: the strings MR-MUF and NCF appear zero times in both texts.