ComfyUI - NVIDIA GPU CUDA Hardware Stratification

Hardware Stratification: Mapping NVIDIA GPUs to ComfyUI & PyTorch

In the current generative AI landscape, the "latest version" is no longer a safe default. The introduction of the Blackwell (RTX 50-series) architecture and the deprecation of Pascal (GTX 10-series) in CUDA 12.8+ have fragmented the ecosystem. This guide provides a precise mapping to align your specific NVIDIA GPU with the correct PyTorch build, ensuring your ComfyUI environment remains functional and avoids the "No kernel image" runtime error.

Why This Matters / The Approach

Neural synthesis performance is dictated by the alignment of silicon and software. To optimize binary sizes, PyTorch maintainers now exclude older architectures from the newest CUDA toolkits.

• Hardware Stratification: Blackwell requires CUDA 12.8+, while Pascal support is removed from those same binaries.
• Compute Capability: Your GPU family defines its SM version (e.g., sm_120 for Blackwell). If the PyTorch binary lacks your SM version, it cannot execute kernels.
• Stability: Understanding your tier prevents "dependency hell" caused by custom nodes forcing incompatible updates.

Prerequisites: Setting Up Your Environment

Baseline requirements are determined by your hardware tier. We must isolate the environment to prevent system-wide driver conflicts.

• NVIDIA Drivers: Version 581.80 or higher is mandatory for RTX 50-series cards.
• Python: Python 3.11 or 3.12 is recommended for the best balance of wheel support and performance.
• Isolation: Use venv to ensure ComfyUI dependencies do not interfere with other projects.

# Verify your driver and current CUDA version
nvidia-smi

Mapping Your Hardware (Architecture Tiers)

1. The Blackwell Tier (Cutting Edge)

• Architecture: Blackwell (Compute Capability sm_120)
• GPUs: RTX 5090, RTX 5080, RTX 5070 Ti, RTX 5070
• ComfyUI Version: Latest Windows Portable or Desktop versions.
• PyTorch Compatibility: PyTorch 2.10+ built with CUDA 12.8 or 13.0.

2. The Mainstream Tier (High Performance)

• Architecture: Ada Lovelace (sm_89) and Ampere (sm_86)
• GPUs: RTX 40-series, RTX 30-series, A-series (A6000, A100)
• ComfyUI Version: Standard Portable or Manual Install.
• PyTorch Compatibility: PyTorch 2.4+ with CUDA 12.1, 12.4, or 12.6.

3. The Legacy Tier (Stability Focus)

• Architecture: Turing (sm_75) and Volta (sm_70)
• GPUs: RTX 20-series, GTX 1660/1650, Titan V, V100
• ComfyUI Version: Manual Install is preferred for granular dependency control.
• PyTorch Compatibility: PyTorch 2.4/2.5 with CUDA 12.1 or 12.4.

4. The Deprecated Tier (Manual Management Required)

• Architecture: Pascal (sm_61) and Maxwell (sm_50/52)
• GPUs: GTX 1080 Ti, 1070, 1060, GTX 900-series
• ComfyUI Version: Avoid any build shipping with CUDA 12.8+.
• PyTorch Compatibility: Must use CUDA 12.6 or 11.8. Higher versions will crash on startup.

Conclusion

This hardware-first approach eliminates the trial-and-error of setting up ComfyUI. The critical distinction is that Pascal users must stay on cu126 or lower, while Blackwell users require cu128+ to even initialize the device. Safety Tip: If a custom node update breaks your environment, immediately re-run the pip install command for your specific tier to restore the correct PyTorch binaries.