Samsung’s HBM4 Delivers 3.3 TB/s Bandwidth, Production Ramps Up

Samsung Electronics has announced mass production of its next-generation HBM4 memory, confirming what it calls the first commercial shipment of HBM4 products to customers. The move comes as memory prices continue to rise and suppliers prioritize short-term supply to meet booming AI demand.

According to details shared via VideoCardz, Samsung’s new HBM4 technology significantly boosts bandwidth, efficiency, and capacity compared to HBM3E.

Samsung HBM4 delivers up to 3.3 TB/s per stack

Samsung says HBM4 reaches a transfer rate of 11.7 Gbps per pin, with tuning support up to 13 Gbps. The company positions 11.7 Gbps as 46% faster than an 8 Gbps industry standard.

Compared to its own HBM3E, Samsung cites 1.22x higher maximum pin speed and up to 3.3 TB/s bandwidth per stack, representing 2.7x higher bandwidth per stack than HBM3E.

The jump in performance comes partly from a major I/O increase. HBM4 doubles pin count from 1,024 to 2,048, enabling higher parallel data throughput for AI accelerators and data center GPUs.

Built on advanced DRAM and logic processes

Samsung’s HBM4 uses its 1c DRAM process, described as a 6th-generation 10nm-class node. The logic base die relies on a 4nm process, further enhancing performance and efficiency.

Initial configurations include 12-layer stacks ranging from 24GB to 36GB, while future 16-layer stacks will scale up to 48GB. These higher capacities target next-gen AI GPUs that require massive on-package memory pools.

Power and thermal gains over HBM3E

Samsung reports major efficiency improvements with HBM4, including 40% better power efficiency than HBM3E, 10% improved thermal resistance, and 30% better heat dissipation.

These gains matter for AI accelerators that run at extreme power levels and operate under sustained workloads in data centers.

AI giants already confirmed as adopters

Several major chipmakers have publicly confirmed HBM4 adoption. NVIDIA will use HBM4 in its Rubin GPU architecture, while AMD lists HBM4 in a 432GB configuration for its Instinct MI430X. Intel is reportedly linked to HBM4 for its next-generation Gaudi platform, codenamed Jaguar Shores.

With AI training clusters scaling rapidly, high-bandwidth memory has become a strategic bottleneck.

Samsung says it is expanding HBM4 production capacity and expects total HBM sales to more than triple in 2026 compared to 2025.

Looking ahead, HBM4E sampling is planned for the second half of 2026, while custom HBM samples are expected to reach customers in 2027.

Meanwhile, other players are exploring alternative memory approaches. AMD and Qualcomm are evaluating SOCAMM2 memory, while some manufacturers are considering Chinese-made memory due to supply shortages and pricing pressure.

As AI hardware demand continues to reshape the semiconductor market, HBM4 appears set to become a core building block for next-generation GPUs and data center accelerators.