Nvidia unveils NVHBM, custom HBM memory with HUGE benefits
Nvidia unveils custom NVHBM memory tech to push bandwidth beyond standard HBM memory
One key limiting factor in modern AI accelerators is memory bandwidth. That’s why every new generation of AI accelerators typically ships with upgraded memory and boosts per-chip memory capacity and bandwidth. Nvidia has unveiled NVHBM, a custom memory type that addresses several issues facing the company’s chip designs. With boosted bandwidth, lower power draw, and lower XPU die requirements, NVHBM can deliver huge benefits over standard HBM memory modules.
Compared to HBM4E, Nvidia claims that NVHBM can deliver 30% more memory bandwidth with 15% less HBM power consumption. By integrating its memory controller into the HBM stack instead of on its connected xPU (CPU, GPU, etc), up to 25% more xPU die space can be allocated to compute, further boosting chip performance.
Nvidia’s clear aim is to move past standardised memory technologies and toward something custom. Nvidia is already working with memory suppliers to make this new memory type a reality.
Right now, it is unknown when Nvidia plans to release products with NVHBM memory. This new memory type isn’t being manufactured, and HBM4E memory has not been used in any commercially available product yet. For context, Nvidia’s new Vera Rubin AI chips use HBM4 memory.
It is likely that Nvidia will use NVHBM memory with next-generation Feynman GPUs. Feynman GPUs are due to launch in 2028. Nvidia could integrate HBM4E into a Vera series GPU refresh (Vera Ultra). Faster memory will be vital for Nvidia’s push for more AI performance. NVHBM could prove to be a critical differentiator for Nvidia’s Feynman GPUs.
You can join the discussion on Nvidia’s NVHBM memory on the OC3D Forums.
