CUDA Compute Capability Reference

What You Can Do on This Page

Check CUDA Compute Capability for RTX 4090, RTX 5090, and All NVIDIA GPUs

CUDA compute capability is essential for determining which GPU features and libraries are available. GPUvec provides official compute capability data for every NVIDIA GPU from the Tesla P100 (compute capability 6.0) and Tesla T4 (7.5) to the RTX 4090 (8.9), RTX 5090 (12.0), and the latest Blackwell architecture. Whether you need to check compute capability for the RTX 2060 (7.5), RTX 4070 (8.9), or A100 (8.0), our comprehensive reference has the official specifications you need.

Find GPU CUDA Compatibility for AI and HPC Frameworks

PyTorch, TensorFlow, and JAX each require specific CUDA compute capability levels. Before renting a cloud GPU instance, verify compatibility by checking our compute capability guide. The RTX 40 series with compute capability 8.9 supports features like FP8 training and improved Transformer Engines. Blackwell GPUs with compute capability 12.0 introduce even more advanced capabilities. Avoid compatibility issues by checking the official compute capability specifications for your GPU model.

Understand NVIDIA GPU Architecture and CUDA Feature Support

Each CUDA compute capability level corresponds to a specific GPU architecture generation. The RTX 30 series (Ampere) has compute capability 8.6 with support for sparse Tensor Cores and FP16/BF16 mixed precision. The RTX 40 series (Ada Lovelace) at 8.9 adds FP8 Transformer Engines. The RTX 50 series (Blackwell) at 12.0 introduces next-generation AI features. Our GPUvec guide explains what each compute capability means for your AI workloads and which features like bfloat16, Tensor Cores, and ray tracing are supported at each level.