NVIDIA GPU architectures

NVIDIA GPU architectures are the microarchitecture generations that NVIDIA ships its chips in. Each generation sets a core design and feature set, such as the Tensor Core generation and the number formats it supports, for cards across several product lines. Blackwell, for example, spans RTX PRO workstation cards and data center racks.

NVIDIA's RTX PRO desktop table groups workstation cards by architecture. The Ampere RTX A6000 and the Ada Lovelace RTX 6000 each carry 48 GB of GDDR6. The Blackwell RTX PRO 6000 carries 96 GB of GDDR7, and the RTX PRO 5500, 5000, 4500, 4000, and 2000 fill out that line. In the data center, Hopper powers the H100 and H200. Blackwell powers the HGX B200 and GB200 NVL72, and Blackwell Ultra powers the HGX B300 and GB300 NVL72. NVIDIA's HGX page also lists HGX Rubin NVL8 and says Vera Rubin is ramping into full production.

Match the architecture to the number format the software needs. Hopper Tensor Cores can mix 8-bit and 16-bit floating point (FP8 and FP16) precision. Blackwell adds new precisions, including 4-bit floating point (FP4). Then confirm the exact part, since one architecture spans a 16 GB workstation card and an eight-GPU HGX board.

Sources

Last verified October 7, 2026.