Nvidia Unveils Custom HBM Design With 30% More Bandwidth Than HBM4E

30% Higher Performance Than HBM4E

Custom Chip Market Poised to Expand

Nvidia has unveiled NVHBM, a next-generation high-bandwidth memory technology. The design is a customized form of HBM that alters the standard configuration made by Samsung Electronics Co. and SK Hynix Inc. to boost the computing power of artificial intelligence accelerators.

Nvidia introduced NVHBM on Aug. 26 alongside its second-quarter earnings report in a post on its blog. The key feature of NVHBM is stronger computing performance. It achieves that by relocating the memory controller in computing chips such as graphics processing units, or GPUs. The memory controller manages which data a GPU pulls from HBM and how it retrieves that data during computation. In the conventional structure, the memory controller sits inside the computing chip. In Nvidia's NVHBM design, it is moved to the base die, the bottom layer of the stacked DRAM in HBM. That frees up space inside the computing chip for additional computing circuitry, lifting performance.

Nvidia said chips built with the NVHBM design would deliver better memory performance and improved power efficiency. Memory bandwidth would rise 30% from HBM4E, the seventh generation of HBM. That means the volume of data transferred each second between HBM and the computing chip would increase by that amount. Power consumption could also fall by as much as 15%, while the area available for computing functions could expand by as much as 25%.

Nvidia is unlikely to manufacture NVHBM itself. The company said it is building a standard NVHBM implementation to be offered through multiple memory suppliers. It also said it would co-develop NVHBM with Annapurna Labs, Amazon's chip design subsidiary. Nvidia is expected to ask memory chipmakers including Samsung Electronics and SK Hynix, which supply it with HBM, to manufacture the product.

The industry expects demand for custom HBM to grow, starting with Nvidia's NVHBM. HBM suppliers such as Samsung Electronics and SK Hynix have so far produced products based on standard specifications. But as AI infrastructure advances, the AI chip specifications sought by customers including big tech companies are diverging. "Starting with Nvidia, customers will gradually demand HBM with different specifications tailored to their own needs," an industry official said.

Samsung Electronics' custom HBM, or cHBM, unveiled at the Hot Chips 2026 semiconductor conference in the US on Aug. 23, was built on the same concept. "Most memory companies see next-generation HBM being developed mainly as custom products," another industry official said.

Roh Yoo-jung, Hankyung.com reporter yjroh@hankyung.com