SM80+ Compute Capability Meaning | Generated by AI

Home 2026.06 English 中文

---

Question: What does NVIDIA SM80+ mean? Are SM versions like Apple’s model numbers (e.g., A1728)?

Answer:

No — it’s the opposite of an arbitrary marketing/model number. SM stands for Streaming Multiprocessor architecture version, and the number encodes the compute capability of the GPU. It’s a structured, ordered version (major.minor), not a SKU identifier like Apple’s A1728.

SM80 = sm_80 = compute capability 8.0. The encoding is SM<major><minor>:

• sm_70 → 7.0 → Volta (V100)
• sm_75 → 7.5 → Turing (RTX 20xx, T4)
• sm_80 → 8.0 → Ampere datacenter (A100)
• sm_86 → 8.6 → Ampere consumer (RTX 30xx, your RTX 4070 is actually sm_89)
• sm_89 → 8.9 → Ada Lovelace (RTX 40xx — your 4070)
• sm_90 → 9.0 → Hopper (H100)
• sm_100 → 10.0 → Blackwell (B100/B200)

Your MI300X is AMD, so it uses gfx942 (CDNA3) instead — a totally separate ISA namespace.

“SM80+” means compute capability ≥ 8.0. It’s a feature gate. Code or kernels marked SM80+ require hardware features introduced in Ampere, most importantly:

• cp.async — asynchronous global→shared memory copies (bypasses registers). This is what makes FlashAttention and modern GEMM pipelines fast.
• 3rd-gen Tensor Cores with bf16, tf32, and async mma instructions.
• Larger shared memory (up to 164 KB/SM on A100).
• Async barriers (mbarrier).

This is why a lot of Triton/CUDA kernels (FlashAttention-2, many fused kernels) say “requires SM80+” — they emit cp.async PTX that simply doesn’t exist on Turing.

The key difference from Apple model numbers: the ordering is semantically meaningful and monotonic. sm_89 > sm_80 tells you Ada is a superset of Ampere features. A1728 vs A1864 tells you nothing — they’re just registry IDs.

Two practical wrinkles:

1. Higher number ≠ strict superset. sm_86/sm_89 (consumer) are newer than sm_80 (A100) but have fewer Tensor Core FLOPs and less shared memory. The minor version sometimes drops datacenter-only features. So “SM80+” guards a feature baseline, not a performance tier.

2. Binary vs PTX. When you compile CUDA you target either a real arch (-arch=sm_89, produces SASS binary) or a virtual arch (-arch=compute_80, produces PTX that JIT-compiles forward to newer GPUs). PyTorch ships fat binaries with TORCH_CUDA_ARCH_LIST="8.0;8.6;8.9;9.0".

Quick check on your boxes:

# Your RTX 4070 → should print 8.9
nvidia-smi --query-gpu=compute_cap --format=csv,noheader

# In Python
python -c "import torch; print(torch.cuda.get_device_capability())"  # (8, 9)

# AMD MI300X — different namespace entirely
rocminfo | grep gfx   # gfx942

So: SM80 is to Apple model numbers as semantic versioning is to a serial number. The number is the capability spec.

---

Back Donate