SM120 Means Compute Capability 12.0 | Generated by AI

Home 2026.07

Question: What does SM120 mean in NVIDIA GPU terminology? Is SM the Streaming Multiprocessor, and do engineers call H200/RTX4070 “SM120”?

Answer:

Your understanding is close, but there are two related meanings:

1. SM = Streaming Multiprocessor

Yes, SM stands for Streaming Multiprocessor. It is the fundamental execution unit inside an NVIDIA GPU.

A GPU is roughly:

GPU
 ├── SM 0
 │    ├── CUDA cores
 │    ├── Tensor cores
 │    ├── registers
 │    ├── shared memory
 │    └── warp schedulers
 │
 ├── SM 1
 │    └── ...
 │
 └── SM N

CUDA schedules work at the SM level. A CUDA kernel launches thread blocks, and thread blocks are assigned to SMs.

(NVIDIA Docs)

2. But “SM120” usually means Compute Capability 12.0

When NVIDIA engineers say:

sm_120
SM120

they usually mean CUDA Compute Capability 12.0, not “the GPU has 120 streaming multiprocessors”.

The naming:

sm_<major><minor>

sm_80  = Compute Capability 8.0
sm_86  = Compute Capability 8.6
sm_89  = Compute Capability 8.9
sm_90  = Compute Capability 9.0
sm_120 = Compute Capability 12.0

CUDA compiler targets are named this way. (NVIDIA Developer)

Example:

nvcc kernel.cu -arch=sm_120

means:

“compile this CUDA kernel optimized for Blackwell compute capability 12.0.”

3. Your examples are mixed

RTX 4070

Ada Lovelace:

Architecture: Ada
Compute Capability: 8.9
CUDA target: sm_89

So:

RTX 4070 -> SM89

not SM120.

H200

Hopper:

Architecture: Hopper
Compute Capability: 9.0
CUDA target: sm_90

So:

H200 -> SM90

not SM120.

RTX 5090 / RTX PRO Blackwell

Blackwell consumer/workstation:

Architecture: Blackwell
Compute Capability: 12.0
CUDA target: sm_120

So:

RTX 5090 -> SM120

(NVIDIA Developer)

The engineer shorthand is basically:

GPU Architecture CUDA target
RTX 4090 Ada sm_89
RTX 4070 Ada sm_89
H100/H200 Hopper sm_90
B200 Blackwell datacenter sm_100
RTX 5090 Blackwell consumer sm_120

(NVIDIA Developer)

For AI kernel development, this matters because libraries like PyTorch, FlashAttention, Triton, CUTLASS, vLLM compile specialized kernels for these targets.

Example:

TORCH_CUDA_ARCH_LIST="8.9;9.0;12.0"

means:

“build kernels for Ada, Hopper, and Blackwell.”

A kernel compiled only for sm_89 may run on newer GPUs through PTX JIT, but it will not necessarily use new hardware features like Blackwell FP4/Tensor Core instructions. (NVIDIA Developer)

So NVIDIA engineers saying “SM120” are speaking the CUDA compiler/hardware capability language, not literally counting SM units.

Back Donate