Nvidia Launches Custom High‑Bandwidth Memory with 30% More Bandwidth and 15% Lower Power Use

• Nvidia’s NVHBM moves the memory controller off the accelerator die and into the HBM stack’s base die, allowing it to claim a potential 30% more bandwidth, 15% lower HBM power, and 25% more compute area versus standard HBM4E
• Nvidia’s NVHBM performance comparisons are linked to HBM4E, which is still in the sampling stage, with a launch expected sometime in 2027
• Access to the custom solution is gated behind NVLink Fusion; Amazon’s Annapurna Labs is the only named partner so far and has not said which GPU uses that memory

On August 26, Nvidia expanded its NVLink Fusion program by introducing NVHBM, a custom high‑bandwidth memory architecture that promises up to 30 % greater bandwidth per stack, a 15 % reduction in HBM power draw, and roughly 25 % more usable area on the accelerator die.

The performance metrics were derived from standard HBM4E samples, and the technology is slated for volume production sometime in 2027.

Amazon’s Annapurna Labs became the first publicly identified partner for Nvidia’s memory advancement, a solution intended to overcome the bandwidth constraints facing frontier‑level AI models.

According to Nvidia’s technical blog, the new architecture reduces interface and support area by up to 67 % compared with the JEDEC HBM4E standard while delivering improvements in power efficiency, memory bandwidth, and usable die area.

Although the concept is not novel—Marvell announced a similar approach in December 2024, partnering with Micron, Samsung and SK hynix and citing up to 25 % more compute area, 33 % higher memory capacity, and a 70 % reduction in memory‑interface power, numbers that are close to Nvidia’s own projections—analyst Neil Shah of Counterpoint Research succinctly noted, “The technology isn’t new; the distribution is.”

While HBM4E has yet to enter mass production—with Samsung delivering its first samples in late May 2026 and SK hynix advancing its own sampling to around June—the NVHBM solution remains even less accessible. It is contingent on NVLink Fusion, Nvidia’s framework for linking third‑party accelerators to its rack‑scale platform, and availability is limited to the program’s approved partners.

Amazon’s Annapurna Labs is the first disclosed participant, and Nvidia’s blog indicates that it will support NVLink Fusion using the Trainium4 accelerator, though it does not directly reference NVHBM.

At present, NVHBM represents a forward‑looking technology that may become widely available only after HBM4E has been broadly deployed. Nvidia’s technical documentation describes NVHBM as employing the same underlying technology planned for future GPUs, without naming the initial generation that will natively support it.

Source link