NVIDIA GPU Compute Capability Levels (Updated Sep 24, 2026)

This is a list of NVIDIA GPU Compute Capability levels.

The various acceleration features supported by inference engines depend on the Compute Capability level, since the hardware accelerators and dedicated functions available differ by level.

Why the numbers do not rise with each generation

Looking at the table, products with a Blackwell GPU are spread across five numbers, 100, 103, 110, 120 and 121.

The number stands for the core GPU architecture. One marketing name can span more than one number when the silicon differs. The split is not by the market a part ships into.

Rubin is 10.7. That is what the CUDA 13.4 release notes state.

Its major number is 10, the same block as B200 and B300. The size of the number does not tell you which generation is newer.

A regular cubin runs only on a GPU with the same major number and an equal or higher minor number. There is no binary compatibility between 100 and 120, even within Blackwell.

For the full explanation see "Blackwell" Was Never One Architecture. Making Sense of NVIDIA's Compute Capability Numbers.

Related Posts

・ PyTorch Versions and Supported NVIDIA GPU Compute Capability Levels

・ 2026 NVIDIA GPU Quick Search Tool

・ 2025 NVIDIA GPU Quick Search Tool

・ "Blackwell" Was Never One Compute Capability. Making Sense of NVIDIA's Compute Capability Numbers