Summary
Per ZDNet and TrendForce reports on October 26-27, Samsung and SK hynix are preparing to mass-produce and ship 8-high stacked HBM4 for Nvidia's Rubin Ultra from the second quarter of 2027 (originally planned with 12-high HBM4E); Nvidia has also quietly cut Vera Rubin's HBM4 bandwidth specification by about 13% per GPU to ease thermal and yield pressure from 12-high stacks and expand qualifying capacity.
Event Overview
- Background: 12-high HBM4 and HBM4E stacks suffer from higher heat generation and lower yields; combined with tight HBM supply, Nvidia discussed new HBM4 supply arrangements with Samsung and SK hynix in Q3 2026.
- Supply shift: Per ZDNet, Samsung and SK hynix plan to mass-produce and ship 8-high HBM4 for Nvidia from Q2 2027; their mainstream high-volume HBM product remains 12-high HBM4. The 8-high products require separate Nvidia qualification.
- Specification cut: Per Wccftech, Nvidia quietly trimmed Vera Rubin's specs starting in September: per-GPU HBM4 bandwidth was cut from 22 TB/s to 19.2 TB/s (about -13%), NVLink 6 throughput from 3.6 TB/s to 3 TB/s (about -17%), and Vera CPU memory was revised from a fixed 1.5 TB of LPDDR5X to "up to 1.5 TB".
Key Information
| Item | Detail |
|---|---|
| Reporting sources | ZDNet (Oct 26), TrendForce (Oct 27), Wccftech |
| Suppliers | Samsung Electronics, SK hynix |
| Customer | Nvidia Rubin Ultra platform |
| Plan change | Original 12-Hi HBM4E plan shifting to 8-Hi HBM4 (possibly mixed with 8-Hi HBM4E) |
| Mass production | From Q2 2027 |
| Reasons | Thermal and yield issues in 12-high stacks; tight HBM supply |
| Rubin GPU bandwidth | Cut from 22 TB/s to 19.2 TB/s (about -13%) |
| NVLink 6 throughput | Cut from 3.6 TB/s to 3 TB/s (about -17%) |
Impact and Market Views
- Cost and capacity: TrendForce says 8-high configurations use fewer DRAM dies, lowering cost per stack and total GPU bill of materials, and enable the same HBM bit output to support more GPUs; amid tight HBM supply, many GPU and ASIC vendors are expected to prioritize 8-high evaluation in 2027.
- Timeline: 12-high HBM4 remains the mainstream configuration in 2026; 8-high adoption is expected to rise sharply from mid-2027.
- Ripple effect: Per Daishin Securities (via Economy Tribune), Anthropic's in-house AI chip could also use 8-Hi HBM4, with Samsung potentially supplying both foundry services and memory.
- Rate data: Samsung and SK hynix disclosed HBM4 data rates of 11.7 Gbps per pin (Samsung cites up to 13 Gbps), and Micron reports speeds exceeding 11 Gbps; these translate to about 3.0 TB/s per stack for Samsung and SK hynix and over 2.8 TB/s for Micron.