alexeigor/sd-benchmarks

Stable Diffusion inference benchmarks

Stable Diffusion XL inference benchmarks

model:

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0

batch size = 1, image size 1024x1024, 50 iterations

A100-SXM, 40GB

Engine	Time
PT2.0,fp16 + compile	5.35 s
Onnxruntime,fp16,ORT_CUDA	4.28 s

RTX 4090, 24GB

Engine	Time
PT2.0,fp16 + compile	6.02 s

RTX 6000 Ada, 48GB

Engine	Time
PT2.0,fp16 + compile	9.07 s

Stable Diffusion inference benchmarks

model:

https://huggingface.co/runwayml/stable-diffusion-v1-5 - sd1.5

https://huggingface.co/stabilityai/stable-diffusion-2-1 - sd2.1

batch size = 1, image size 512x512, 50 iterations

A100-SXM, 40GB

Engine	Time, sd1.5	Time, sd2.1, 512x512	Time, sd2.1, 768x768
PT2.0,fp16	1.96 s (4.54gb VRAM)
PT2.0,fp16 + compile	1.36 s (5.96 gb)		2.37 s
AITemplate,fp16	1.01 s (4.06 gb)
DeepSpeed,fp16	1.18 s		2.28 s
Oneflow,fp16	0.98 s (5.62 gb)
TensorRT 8.6.1, fp16	0.98 s	0.81 s	1.88 s
TensorRT 10.0, fp16			1.57 s
Onnxruntime,fp16,ORT_CUDA	0.85 s	0.76 s	1.63 s
Jax,XLA,bf16	1.58 s	1.35 s	3.61 s

H100-PCIe, 80GB

Engine	Time, sd1.5	Time, sd2.1
PT2.0,fp16	1.44 s
PT2.0,fp16,compile	1.11 s
TensorRT 8.6.1,fp16	0.75 s	0.68 s
Jax,XLA,bf16	1.18 s

H100-SXM, 80GB

Engine	Time, sd1.5	Time, sd2.1	Time, sd2.1, 768x768
PT2.0,fp16,compile	0.83 s	0.70 s	1.39 s
TensorRT 8.6.1,fp16	0.49 s	0.48 s	1.05 s
Jax,XLA,bf16	1.00 s	0.79 s

RTX 4090, 24GB

Engine	Time, sd1.5	Time, sd2.1	Time, sd2.1, 768x768
PT2.0,fp16,compile	1.17 s		2.26 s
TensorRT 8.6.1, fp16	0.74 s	0.68 s	1.52 s

L40, 48GB

Engine	Time, sd1.5	Time, sd2.1	Time, sd2.1, 768x768
PT2.0,fp16,compile	2.09 s		3.08 s
TensorRT 8.6.2,fp16	0.91 s		2.19 s

RTX 6000 Ada, 48GB

Engine	Time, sd1.5	Time, sd2.1	Time, sd2.1, 768x768
PT2.0,fp16,compile	1.28 s		2.77 s
TensorRT 8.6.2,fp16	0.90 s		2.25 s

V100, T4

GPU	PT2.0,fp16,xformers
V100, 16gb	2.96 s
T4, 16gb	7.83 s

How to run

Ubuntu, Debian VM setup https://gist.github.com/alexeigor/b4c21b5e1fe62d670c433d4ac8c9fd83

docker build -f ./Dockerfile --network=host --build-arg HF_TOKEN=xxxxx -t test_pt .

docker run -it --network=host -v ${PWD}/workspace:/workspace -w /workspace --gpus all --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 test_engine

References:

• https://github.com/microsoft/DeepSpeed-MII/tree/main/examples/benchmark/txt2img
• https://github.com/Oneflow-Inc/oneflow
• https://github.com/Oneflow-Inc/diffusers
• https://github.com/facebookincubator/AITemplate
• https://arxiv.org/abs/2304.11267
• https://github.com/dbolya/tomesd
• https://huggingface.co/docs/diffusers/main/en/optimization/fp16
• https://github.com/hidet-org/hidet
• https://github.com/stochasticai/x-stable-diffusion
• https://github.com/microsoft/onnxruntime/tree/main/onnxruntime/python/tools/transformers/models/stable_diffusion
• https://github.com/microsoft/Olive
• https://github.com/NVIDIA/TensorRT/tree/main/demo/Diffusion (kudos to Denis Timonin https://www.linkedin.com/in/denistimonin/)
• https://medium.com/microsoftazure/accelerating-stable-diffusion-inference-with-onnx-runtime-203bd7728540
• https://huggingface.co/blog/sdxl_jax
• https://huggingface.co/blog/simple_sdxl_optimizations
• https://pytorch.org/blog/accelerating-generative-ai-3/