Specifications
Product
Micron HBM4 12-high (12-H)
Capacity
36 GB per stack
Stack Height
12-high
Interface
2,048-bit (twice HBM3E)
Pin Speed
>11 Gb/s
Bandwidth per Stack
>2.8 TB/s
Bandwidth Gain
≈2.3× Micron HBM3E 12-high
Power Efficiency
>20 % better pJ/bit than HBM3E at similar speeds
Channels / Pseudo-channels
32 / 64 (doubled parallelism)
Banks per Die
256 (double HBM3E's 128)
DRAM Process
Micron 1β (1-beta)
Base Die
Logic base die from TSMC (12FFC+ / N5 class)
First Platform
NVIDIA Vera Rubin
Production Status
High-volume production since Q1 2026
Next Density
48 GB 16-high sampling to customers
Successor
HBM4E — expected 2027, TSMC base die + 1γ DRAM
Role
AI / HPC accelerators — not a DIMM replacement
Compliance
JEDEC HBM4
Lead Time
Allocation based — contact for availability
MOQ
Contact for allocation
Overview
Micron's 36 GB 12-high HBM4 is the shipping 2026 generation of high-bandwidth memory for AI and HPC accelerators. It doubles the bus width to 2,048 bits and pushes pin speed past 11 Gb/s, producing more than 2.8 TB/s of bandwidth from a single stack — roughly 2.3× the equivalent 12-high HBM3E.
The efficiency gain is as important as the bandwidth: more than 20 % better pJ/bit means the same data movement costs measurably less power, which is what makes sustained inference at 1,000 W+ accelerator power budgets thermally feasible.
Micron reached high-volume production in the first quarter of 2026, aligned to NVIDIA's Vera Rubin platform, with 48 GB 16-high samples following for higher-capacity accelerators.
Key Benefits
2,048-bit interface — double HBM3E bus width; >2.8 TB/s per stack at >11 Gb/s pin speed; >20 % better pJ/bit for lower thermal cost at the same bandwidth; in volume production for NVIDIA Vera Rubin rather than sampling.
Applications
HBM4 is the memory tier for AI training and inference accelerators, HPC vector engines and any ASIC where model weights must sit within a few hundred nanoseconds of the compute die. It is not a substitute for DDR5 RDIMM capacity.
Request a Quote — MICRON HBM4 36 GB 12-HI — 2,048-BIT, >2.8 TB/S STACK MEMORY
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →