Published November 16, 2026
Where the energy of a one-step image goes
sd-turbo makes a 512x512 image in a single denoising step. Measured stage by stage on my RTX 3080 Ti with the NVML energy counter, the step itself is not where most of the energy goes: at one step, the VAE decoder that turns the result into pixels takes 62% of the energy above idle and more than half of the time.
Distilled diffusion models such as sd-turbo are sold on their step count: one denoising step instead of twenty or fifty, so an image in about a tenth of a second. The step count is the number everyone tunes. I wanted to know where the time and the energy of one image actually go, stage by stage, with the measurement rules from the previous post, on the same RTX 3080 Ti.
The pipeline is plain diffusers in eager mode: sd-turbo at 512×512, fp16, batch 1, with 1, 2 and 4 steps, 200 measured images for each after 10 warm-up ones. An image goes through three stages: the text encoder turns the prompt into embeddings, the UNet denoises a latent once per step, and the VAE decoder turns the final latent into pixels. CUDA events around each stage give its time.
Why the energy cannot be read per image
An image takes 112 to 211 ms. On this card the energy counter moves every 100 ms and under-counts when read faster than about twice a second: by 14% at 10 Hz. Reading it around each image, let alone each stage, would measure the counter rather than the pipeline.
So each stage is replayed alone, on inputs captured from a real image, in a loop of about 3 s: about 245 text encoder calls, 90 UNet calls or 49 VAE decodes per block. The whole pipeline is run the same way, 14 to 27 images per block. The four kinds of block alternate, 15 rounds per step count, and the counter is read once per block: the reading that ends one block starts the next. Energy per call is the block energy divided by the number of calls. Dynamic energy subtracts idle power, 108 W measured over 30 s, times the block duration.
The result
| steps | ms per image | J per image | J above idle | text encoder | UNet | VAE decode | other |
|---|---|---|---|---|---|---|---|
| 1 | 111.9 | 35.8 | 23.7 | 3% | 29% | 62% | 6% |
| 2 | 145.6 | 45.6 | 30.2 | 3% | 46% | 48% | 3% |
| 4 | 211.1 | 66.9 | 44.0 | 2% | 63% | 32% | 3% |
Per call, the text encoder takes 12.3 ms and 0.81 J above idle, one UNet step 33 ms and 6.9 J, the VAE decode 61 ms and 14.6 J. "Other" is the whole pipeline minus the three stages: the scheduler, the kernel launches and the gaps between kernels. It comes out at 3 to 6%, which is the check that matters for the method: the stages measured alone add up to the pipeline measured as a whole.
The decoder, not the denoiser
The UNet is the part that scales with the step count, and it behaves exactly as expected: each step adds 33 ms and 6.9 J. Going from four steps to one cuts the time per image by 47% and the energy above idle by 46%. But the VAE decode does not move. It runs once per image whatever the step count, so as the UNet shrinks, the decoder’s share grows: 32% of the energy at four steps, 62% at one. At one step it is also 61 of the 112 ms, more than half of the time.
The mean power of the pipeline barely changes with the step count, 320 to 322 W, even though the stages draw very differently when run alone: 172 W for the text encoder, 340 W for the VAE decode. Across these three configurations, energy per image follows time per image almost exactly. That will not hold in general, as the DBSCAN study showed, but it means that here, the time profile and the energy profile point at the same culprit.
For a real-time loop at one step, then, the lever is no longer the step count. It is the decoder: a smaller or distilled autoencoder, or a compiled one. I have not measured any of those yet; that is the obvious next run. The text encoder is a smaller lever: a loop with a fixed prompt could cache its output and save 2 to 3%.
What this does not show
- The counter has not been checked against an external meter. In the probes it agreed with the NVML power value within 5% when read once or twice a second; here it is read every 3 s. NVML also covers the board only: the CPU, which runs the scheduler and the kernel launches, is not counted.
- Idle is a desktop session with the display on and the model in VRAM. The GPU did not drop to a low-power state: at idle the SM clock read 1770 MHz. Dynamic energy is what the work adds on top of that state, not on top of a GPU at rest.
- A stage replayed alone runs back to back, without the gaps it has inside the pipeline. That "other" comes out at only 3 to 6% suggests the difference is small.
- During the loaded blocks the SM clock held between 1920 and 2025 MHz and the GPU warmed from 41 to 60 °C over the one-step blocks, then stayed at 60 to 62 °C. NVML reported a software power cap after every one of these blocks, although the mean power stayed under the 370 W limit.
- Plain diffusers, no
torch.compile, CUDA graphs or TensorRT. One GPU, one driver (610.88), Windows 11, one prompt, one run.
The measurement script, the raw blocks and images, and the script that rebuilds the table and the figure are on GitHub: gpu-sd-study. A run takes about twelve minutes on a CUDA GPU.