Micron roadmap: HBM4 at 2.8 TB/s per stack, vertical HBM on GPUs by 2028
rohanpaul_ai · x · 2026-10-24
Micron's latest report details its memory and storage roadmap through 2030:
- HBM4 bandwidth: a single stack moves 2.8 TB/s; 13 flagship AI GPUs (8 stacks each) match the world's entire international internet capacity of 282 TB/s.
- Exploding memory demand: per-GPU memory grew 3.6x from H100's 80GB to B300's 288GB; Micron estimates AI context length grows 30x per year; a robotaxi needs over 10x the memory of today's driver-assist car.
- Vertical HBM: planned for 2028, stacking memory directly on top of the GPU to shorten data paths and improve performance per watt.
- AI efficiency gains: agents writing chip logic and firmware cut SSD controller design time by 50% and customer-ready time by 60%.
- Dropping 4F² DRAM: after 20+ years, Micron abandons the denser cell layout as too costly, splitting from Samsung and SK hynix, and will build DRAM upward in 3D instead.
- AI-class NAND: flash GPUs can read directly in small chunks, plus HBM-style stacked flash, targeting 25M-100M ops/sec for model weights and long conversations.
- Density: over 1,500 memory capacitors fit across a 50µm hair; each DRAM wafer holds 400x more bits than in 1995.
More from Infra
- antirez makes the case for local AI as hacker's Samsung TV snooping saga hits HN — antirez · 2026-10-24
- PNG is slow because of deflate, not libpng — zlib-ng's SIMD wins — ssh4net · 2026-10-24
- Hunyuan 3D 2.1 nearly unusable on Mac: errors above 512 resolution — idmimagineering · 2026-10-24
- GPU programming day 293: diving into SGLang via mini-sglang and MIT's EfficientML lecture 12 — NandoDF · 2026-10-24
- China's Jiuzhang 4.0 claims 10^42-year speedup in 25 microseconds — but the task has no practical use — AryHHAry · 2026-10-24
- Fancy coding-agent harnesses cost 3x more but don't improve pass rates, study finds — The Decoder · 2026-10-24