THE CRUNCH
Samsung's upcoming HBM4E memory has reportedly passed testing for NVIDIA's next-generation "Rubin Ultra" AI accelerator family, according to a report by Hankyung picked up by TechPowerUp. It is a notable turnaround for Samsung, whose previous HBM3E did not meet NVIDIA's requirements, leaving the company supplying only GDDR7 to NVIDIA instead. The new memory runs at 14 Gbps per pin with 2,048 pins per stack, giving a maximum bandwidth of 3.6 TB/s per stack, rising to 4 TB/s at 16 Gbps.
Samsung is currently offering HBM4E in 12-layer stacks of 48 GB, with 8-layer 32 GB and 16-layer 64 GB configurations expected to follow customer demand. Samples have been going out since May, and mass production is expected to reach around 20,000 wafers per month in 2026, doubling to 50,000 wafers per month in 2027.
For the base die, the layer that handles the stack's logic, Samsung has reportedly allocated half of its 4 nm SF4 node capacity to HBM4E base die manufacturing. The report suggests this points to high customer interest in customised memory designs such as NVHBM, NVIDIA's approach to custom HBM. With roughly a dozen memory stacks per AI accelerator, total bandwidth per GPU would be substantial.