Which processes run faster on the graphics card, and by how much
A modern graphics card does in a blink what a processor takes seconds to do. DSPX puts it to work for sharpening, noise reduction, stretching, geometry, layers and the heaviest steps of pre-processing, and asks for nothing: with the default setting the program uses the card when it pays and the processor when that is faster. This chapter lists what gains, by roughly how much, what to expect, and where the choice is made.
The figures are typical for a recent discrete graphics card and images of several megapixels, compared with the same process on the processor of the same computer. They vary with the card and the image; we round down, and "up to" means the best case we measured.
| Process | Typical gain | Notes |
|---|---|---|
| Sharpening and detail | ||
| Deconvolution | up to 30x | A 24 MP color image at the default settings: from minutes to seconds. |
| Texture Sharpen | up to 40x | |
| Multiscale Sharpen | about 15x | |
| Clarity | about 10x | |
| Flow Sharpen | about 10x | |
| Wavelets | 8x to 12x | All targets, Lightness included. |
| HDR Compression | 8x to 12x | All targets, Lightness included. |
| Filament Sharpen | 4x to 6x | |
| Unsharp Mask | 1.5x to 4x | The least, because it is quick on the processor too. |
| Blur | about 3x, up to 8x | The larger the filter, the larger the gain. A small filter stays on the processor unless the image is very large. |
| Deep Sharpen, Deep Star Removal, Deep Star Mask | Accelerated | The networks run on the graphics card on their own; see The AI Processes. |
| Noise | ||
| Noise Reduction | about 4x | The whole process, on a 24 MP color image or a 60 MP mono one. |
| Multiscale TGV Denoise | about 4x | The solver, the largest part of the run, on a 24 MP gray image; the same result to within 5 thousandths of a noise sigma. |
| Stretch and color | ||
| Color Calibration | about 10x | For the part that finds the structures; the rest is as quick on the processor. |
| Masked Stretch | about 10x | Live preview updates. |
| Exponential Transformation | about 3x | Live preview updates. |
| Saturation, SCNR | 3x to 5x | Live preview updates. |
| Hyperbolic Stretch, Histogram | 2x to 3x | Live preview updates. |
| Curves | about 2x | |
| Background Neutralization | about 2x | On images of about 10 MP and more. |
| Adaptive Stretch, Arcsinh Stretch | accelerated | Part of the work runs on the card; we give no figure. |
| Several of these steps in one container | about 6x | The image stays on the card from one step to the next; the more steps, the more it gains. |
| Geometry | ||
| Resample, Rotate, a turned Crop, Dynamic Astro Crop | 2.5x to 4x | With the high-quality interpolations, which have the most to compute. The simple ones stay on the processor, which is as quick there. |
| Surface Align | 2x to 3x | Putting the frames together on a canvas of 24 to 60 MP: the resampling of every frame and, with multi-band blending, the whole blend (1.5x to 2x when the blend is a plain feather). The search for the match between the frames and the seams stay on the processor, so a whole run gains less. |
| Layers | ||
| Moving an opacity or blending slider on a stack of layers | about 7x | Large canvases of several layers, on a discrete card. |
| Previewing a process on a layer | up to 2x | At the larger zoom reductions; about even at the smaller ones. |
| Pre-processing | ||
| Registration | 1.4x to 2x | Aligning the frames of a large set; the more modest the processor, the larger the gain. |
| Integration | up to 1.3x | On a processor with few cores. The program picks the faster of the two for the computer it finds. |
| Previews | ||
| Live preview of Wavelets, HDR Compression and Unsharp Mask | 4x to 13x | On a 24 MP image. The panel answers while the slider moves. |
Everything else, which includes Calibration, Drizzle, Local Normalization, the mask processes, PixelMath and the file formats, runs on the processor. Invert and Binarize use the card only inside a container with other steps that do; alone they are quicker on the processor.
The graphics card is on by default. The choice is one setting, in , and it holds for the AI processes and for everything in the table.
Processor: Automatic uses the best graphics card there is, for each step that gains from it, and the processor for the rest. CPU only never uses the card. A card named in the list is used by every step in the table except the AI processes, which keep Automatic. A line beside the list says what the program found.
Test the GPU: Times a short job on a graphics chip built into the processor and on the processor, and remembers the verdict. Automatic uses such a chip only if it was faster. A card chosen by name is used either way.
Every run says where it ran: the done line of the process ends in "(GPU)" or "(CPU)", the step in the History panel does too, and the tooltip of a step that stayed on the processor says why, for example that the image is small.
The AI processes follow the same rule: a separate card, or the chip of a Mac with Apple silicon, runs the networks; a built-in chip is slower than the processor for them.
The two kinds of work reach the card differently. The AI processes' networks go through DirectML on Windows, Core ML on Apple silicon, or CUDA where ONNX Runtime has a CUDA build. Every other step in the table goes through Vulkan, which the graphics driver provides on Windows and Linux; on a Mac, DSPX carries its own copy that runs on Metal. A card whose driver has no Vulkan, as with the basic Windows driver, leaves those steps to the processor.