Nvidia introduced NVHBM in late August 2026, a custom HBM variant that moves the memory controller from the main chip into the memory stack's base die. This extension of the NVLink Fusion platform provides partners with a custom base die and modified PHY.
The technology targets hyperscalers like Amazon's Annapurna Labs, the first partner to use NVHBM for its Trainium4 accelerators. While GeForce cards still use GDDR7, Nvidia claims NVHBM offers 30% more bandwidth and 15% lower power than standard HBM4E.
By shifting logic to the base die, NVHBM reduces the PHY area by 67% and saves up to 30% of space on the computing chip. This implementation leaves a narrower interface on the main chip, utilizing technology Nvidia plans to integrate into its future GPUs.