Micron, Samsung, SK hynix HBM Roadmaps: HBM4 and Beyond
▼ Summary
– High Bandwidth Memory (HBM) is critical for AI systems, providing unmatched data speeds to power advanced GPUs and accelerators.
– Major DRAM makers (Micron, Samsung, SK hynix) are producing 8-Hi HBM3E and developing 12-Hi HBM3E, with SK hynix leading mass production.
– HBM4 and HBM4E are in development, featuring a 2048-bit interface, higher speeds, and up to 16-Hi stacks, but initial versions will use 24 Gb dies.
– HBM4 production is expected to start in 2026, with HBM4E following in 2027, using advanced base dies from TSMC or Samsung Foundry.
– Despite shortages, HBM remains the standard for AI and HPC due to its superior bandwidth, with ongoing innovations to meet growing demand.
High Bandwidth Memory (HBM) has become the backbone of modern AI and high-performance computing, delivering the massive data throughput required by cutting-edge GPUs and accelerators. As demand for faster, more efficient memory grows, industry leaders Micron, Samsung, and SK hynix are pushing the boundaries with next-generation HBM technologies. The race is on to develop higher-capacity, faster, and more power-efficient solutions that will fuel the next wave of AI advancements.
Currently, all three manufacturers are producing 8-Hi HBM3E stacks, the latest iteration of this critical memory technology. However, 12-Hi HBM3E is already on the horizon, promising even greater capacity and performance. Meanwhile, HBM4 and HBM4E loom in the distance, set to redefine memory bandwidth with wider interfaces and higher layer counts.
The Need for Speed: Why HBM Dominates AI Workloads
This leap in performance comes from stacking multiple DRAM dies vertically using through-silicon vias (TSVs), allowing for dense, high-speed interconnects. The number of stacked layers, denoted as 8-Hi, 12-Hi, or even 16-Hi, directly impacts capacity and efficiency. Given these advantages, HBM remains the gold standard for AI accelerators, HPC chips, and high-end GPUs.
12-Hi HBM3E: The Next Big Leap
SK hynix has already begun mass production of 12-Hi HBM3E, while Micron is sampling its version ahead of full-scale manufacturing. Samsung, however, has faced delays, likely due to its reliance on older 1α process technology, while competitors use more advanced 1ß (5th Gen 10nm-class) nodes. Still, Samsung is expected to catch up in time for B300 production.
HBM4 & HBM4E: The Future of Memory
HBM4E, meanwhile, is expected to push speeds beyond 9 GT/s, surpassing HBM4’s standard 6.4 GT/s. Memory makers may also integrate custom base dies with additional features like enhanced caches or specialized interfaces.
Manufacturer Roadmaps: Who Leads the Pack?
Production timelines suggest HBM4 samples arriving in late 2025, with mass production in 2026. HBM4E is expected in late 2027, aligning with next-gen AI accelerators like Nvidia’s Rubin and AMD’s MI400.
(Source: Tom’s Hardware)
