Nvidia moves the memory controller into the HBM base die, with Annapurna Labs first in line
NVIDIA AI Blog·medium signal
NVIDIA announced NVHBM on August 26, relocating its custom memory controller from the compute chip into the HBM base die. It claims up to 30% more memory bandwidth than standard HBM4E, 15% lower HBM power, and up to 25% more freed area on the XPU compute die. Amazon's Annapurna Labs is the first partner, with next-generation Trainium4 supporting NVHBM alongside NVLink Fusion, which is the more interesting signal: Nvidia is selling memory architecture into the very custom silicon that competes with its GPUs.