Understanding NVIDIA’s sm_80 to sm_120 in Plain English

If you’ve worked with CUDA, PyTorch, TensorRT, or AI tools like ComfyUI, you’ve probably seen labels like sm_80, sm_86, sm_90, or sm_120.

For the longest time, these numbers popped up in ComfyUI nodes, and I honestly didn’t know what they meant—until recently. After digging into it, I realized the explanation is actually pretty simple.

What “SM” Means

SM stands for Streaming Multiprocessor, the core building block inside an NVIDIA GPU.
When software refers to sm_80 or sm_120, it’s talking about the GPU’s compute capability—basically the GPU’s feature level.

Think of compute capability like a version number on a game console.
Higher number → newer features, more instructions, better performance.

Breaking Down the SM Versions (sm_80 → sm_120)

sm_80 – Ampere (A100, workstation/server variants)

Released around 2020. Brought improvements to Tensor Cores and FP16 performance.

sm_86 – Ampere Consumer (RTX 30 Series)

3080, 3090, 3070, 3060, etc.
Slightly different features from sm_80—this is why some model wheels don’t work across both.

sm_89 – Ada (Laptop/Workstation Variants)

Used in mobile 4090/4080 and Ada workstation cards.

sm_90 – Hopper (H100)

A massive jump for AI workloads:

• FP8 support

• Transformer Engine

• Huge improvements for LLMs

sm_100 – Blackwell (Data Center)

Next-generation AI architecture.
Designed for extremely large-scale training.

sm_120 – Blackwell Consumer (RTX 50 Series)

This is the one that confused many people (including me).
sm_120 is used by the GeForce RTX 5080/5090 and other upcoming 50-series GPUs.
It’s Blackwell architecture tuned for consumer cards.

Why These Numbers Matter

Most users only notice SM versions when something breaks.
For example:

• A PyTorch wheel doesn’t support your GPU

• CUDA tools throw “unsupported architecture sm_XX” errors

• A ComfyUI node fails to compile kernels

• A model requires a higher compute capability than your card supports

A few common situations:

• RTX 30 Series (sm_86) → won’t run binaries built only for sm_90

• RTX 50 Series (sm_120) → may need updated CUDA versions for compatibility

• Older builds often don’t include sm_120 yet, causing early driver/tool issues

Knowing your GPU’s SM version makes troubleshooting much easier.

Quick Reference Table

The Bottom Line

I used to think sm_80, sm_90, and so on were just internal GPU codes.
But after looking into it recently, I finally understood:

• They’re version numbers for GPU features

• They decide which software builds your GPU can run

• They determine CUDA compatibility

• They explain why some models run on one GPU but fail on another

Once you understand SM versions, troubleshooting AI tools becomes much easier—especially with newer architectures like sm_120 entering the scene.

NVIDIA Official Docs

• Compute Capability Overview
  https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#compute-capabilities
  (Explains SM versions and architecture mapping.)

• List of All CUDA GPUs and Their SM Versions
  https://developer.nvidia.com/cuda-gpus
  (A clean chart showing all GPU models and their compute capability.)

• CUDA Release Notes
  https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html
  (Useful when new GPUs like sm_120 arrive and older CUDA versions don’t support them.)

Developer and Open-Source References

• PyTorch Get Started (GPU Compatibility)
  https://pytorch.org/get-started/locally/
  (Explains which CUDA builds support which GPUs.)

• TensorRT Developer Guide
  https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html
  (Good for readers doing model optimization on new architectures.)

• Hugging Face Hardware Compatibility Notes
  https://huggingface.co/docs
  (Helps readers understand why some models fail to load on older GPUs.)

Further Reading

How to Install Triton and SageAttention on Windows

How to Upgrade FramePack Studio for NVIDIA 50-Series GPUs