RGB Input Pipeline Benchmark
Compare RGB JPEG-to-CUDA throughput and peak GPU memory across AlbumentationsX, Pillow, TorchVision, Kornia, and NVIDIA DALI. 73 recipes, 8 execution paths.
RGB input pipeline results
Every path reads RGB JPEGs, prepares the recipe, and delivers a synchronized CUDA batch. CPU and GPU labels identify where most augmentation stages run. Every path normalizes on GPU.
Each comparison uses its own shared recipe set. Compare paths within each chart. The table below includes every measured recipe for these paths, including recipes outside the summary set. Recipe names are shortened; hover over a name for its full pipeline.
Higher throughput is better. Values are medians across seeds. Hover for the observed range. A dash means no measured result.
| R01Resize224 | 4,691 | 2,665 | 3,854 | 3,425 | 2,329 | 2,373 | 5,056 | 4,176 |
|---|---|---|---|---|---|---|---|---|
| R02RandomCrop224 | 4,421 | 4,319 | 4,120 | 4,464 | 2,256 | 2,255 | 6,387 | 4,197 |
| R03RandomResizedCrop | 4,706 | 2,714 | 3,587 | 2,336 | 1,923 | 1,955 | 6,065 | 3,904 |
| R04HorizontalFlip | 3,955 | 4,287 | 4,179 | 3,645 | 2,128 | 2,203 | 6,417 | 4,291 |
| R05VerticalFlip | 4,279 | 4,255 | 4,220 | 3,495 | 2,073 | 2,207 | 6,401 | 4,146 |
| R06Pad+RandomCrop224 | 4,244 | 3,891 | 3,547 | 3,147 | 1,792 | 1,848 | 4,996 | 4,022 |
| R07Rotate | 4,409 | 4,057 | 2,499 | 1,245 | 1,520 | 2,248 | 6,379 | 4,192 |
| R08Affine | 4,106 | 2,625 | 2,652 | 1,280 | 1,533 | 2,271 | 6,433 | 3,902 |
| R09Perspective | 4,180 | 2,650 | 2,645 | 1,262 | 1,536 | 2,306 | 6,127 | 3,669 |
| R10Elastic | 3,032 | — | 236 | 22 | 793 | 2,371 | — | — |
| R11ColorJitter | 3,569 | — | 967 | 689 | 968 | 331 | 3,511 | — |
| R12ChannelShuffle | 4,751 | 4,184 | 3,986 | 2,867 | 2,113 | 2,158 | 6,251 | 4,192 |
| R13Grayscale | 4,804 | 4,104 | 3,841 | 3,152 | 2,021 | 2,197 | 6,416 | 4,100 |
| R14RGBShift | 4,655 | 3,833 | — | — | 2,028 | 2,259 | 6,190 | 4,006 |
| R15GaussianBlur | 4,399 | 2,521 | 2,168 | 1,904 | 1,226 | 2,134 | 6,329 | 3,764 |
| R16GaussianNoise | 3,376 | — | 1,231 | 3,132 | 1,745 | 2,213 | 6,372 | — |
| R17Invert | 4,928 | 3,940 | 4,221 | 3,483 | 2,070 | 2,182 | 6,392 | 4,186 |
| R18Posterize | 4,968 | 3,966 | 4,147 | 3,075 | 1,662 | 2,225 | 6,371 | 4,029 |
| R19Solarize | 4,622 | 3,820 | 3,583 | 3,168 | 1,911 | 2,212 | 6,401 | 4,245 |
| R20Sharpen | 4,481 | — | 2,307 | 1,941 | 1,282 | 2,264 | — | — |
| R21AutoContrast | 4,160 | 3,724 | 2,718 | 2,344 | 1,834 | 2,158 | — | 3,882 |
| R22Equalize | 3,976 | 3,750 | 2,973 | 1,293 | 1,514 | 2,219 | 6,388 | 3,872 |
| R23Erasing | 4,726 | — | 3,877 | 1,932 | 1,761 | 2,123 | 6,293 | — |
| R24JpegCompression | 3,868 | 3,511 | 3,328 | 2,360 | 668 | 2,086 | 6,340 | 3,611 |
| R25RandomGamma | 4,586 | 4,054 | 3,059 | 3,821 | 1,715 | 2,247 | 6,427 | 4,095 |
| R26PlankianJitter | 4,340 | — | — | — | 1,972 | 2,225 | — | — |
| R27MedianBlur | 3,758 | 243 | — | — | 405 | 762 | 5,080 | 3,916 |
| R28MotionBlur | 4,213 | — | — | — | 1,474 | 2,241 | — | — |
| R29CLAHE | 3,109 | — | — | — | 812 | 2,162 | 5,955 | — |
| R30Brightness | 4,633 | 3,954 | 3,883 | 3,562 | 2,055 | 2,151 | 6,344 | 3,866 |
| R31Contrast | 4,380 | 3,690 | 3,428 | 2,676 | 1,965 | 2,178 | 6,300 | 2,439 |
| R32Blur | 4,851 | 3,341 | — | — | 1,058 | 2,211 | 6,349 | 4,141 |
| R33ChannelDropout | 4,732 | 4,072 | — | — | 2,059 | 2,180 | 6,289 | 4,264 |
| R34LinearIllumination | 3,826 | — | — | — | 1,816 | 2,174 | — | — |
| R35CornerIllumination | 3,991 | — | — | — | 1,587 | 2,115 | — | — |
| R36GaussianIllumination | 3,789 | — | — | — | 1,527 | 2,206 | — | — |
| R37Hue | 4,064 | — | 1,403 | 838 | 1,089 | 2,191 | 6,357 | — |
| R38PlasmaBrightness | 2,603 | — | — | — | 481 | 2,127 | — | — |
| R39PlasmaContrast | 2,202 | — | — | — | 472 | 2,154 | — | — |
| R40PlasmaShadow | 2,502 | — | — | — | 797 | 2,225 | — | — |
| R41Rain | 3,979 | — | — | — | 1,541 | 2,154 | — | — |
| R42SaltAndPepper | 4,750 | — | — | — | 1,533 | 800 | 6,332 | — |
| R43Saturation | 4,354 | 3,798 | — | — | 1,975 | 2,363 | 6,323 | 4,030 |
| R44Snow | 3,733 | — | — | — | 1,085 | 2,227 | — | — |
| R45OpticalDistortion | 3,817 | — | — | — | 1,303 | 2,306 | — | — |
| R46Shear | 3,756 | 2,643 | 2,839 | 1,526 | 1,554 | 2,274 | 6,436 | 3,822 |
| R47ThinPlateSpline | 3,139 | — | — | — | 643 | 2,306 | — | — |
| R48PhotoMetricDistort | 3,482 | — | 1,172 | 635 | — | — | — | — |
| R49ColorJiggle | 3,640 | — | 1,210 | 687 | 975 | 333 | 3,461 | — |
| R50LongestMaxSize+RandomCrop224 | 3,729 | 1,236 | 3,380 | 2,492 | 1,263 | 1,285 | 5,023 | 2,950 |
| R51SmallestMaxSize+RandomCrop224 | 3,187 | 848 | 3,079 | 2,593 | 980 | 971 | 5,025 | 2,403 |
| R52Transpose | 4,945 | 4,158 | 4,014 | 4,168 | 1,687 | 2,130 | 6,383 | 4,342 |
| R53RandomRotate90 | 4,846 | 4,247 | 4,149 | 3,840 | 1,389 | 2,204 | 6,383 | 4,368 |
| R54RandomJigsaw | 4,442 | — | — | — | 1,878 | 2,245 | — | — |
| R55EnhanceEdge | 4,480 | 2,699 | — | — | — | — | — | 3,995 |
| R56EnhanceDetail | 4,412 | 2,762 | — | — | — | — | — | 4,046 |
| R57UnsharpMask | 4,252 | 2,276 | — | — | — | — | — | 3,604 |
| R58CenterCrop | 4,658 | 4,206 | 4,240 | 5,997 | 2,744 | 2,663 | 6,506 | 4,358 |
| R59Resize256+Crop | 4,595 | 2,342 | 3,605 | 3,302 | 2,011 | 2,092 | 5,091 | 3,805 |
| R60Resize224+CropAndPad | 4,284 | 2,673 | 2,900 | 2,835 | 2,210 | 2,264 | 5,066 | 3,979 |
| R61LetterBox | 4,649 | 2,876 | 3,772 | 2,762 | 2,366 | 2,593 | 5,054 | 3,987 |
| R62PadIfNeeded+RandomCrop224 | 4,404 | 4,123 | 4,275 | 4,014 | 2,328 | 2,173 | 5,007 | 4,400 |
| R63Resize224+ChannelSwap | 4,569 | 2,657 | 3,585 | 2,636 | 2,130 | 2,266 | 5,063 | 3,722 |
| R64Resize224+Colorize | 4,267 | 2,409 | — | — | 1,417 | 2,253 | 5,058 | 3,531 |
| R65Resize224+Dithering | 895 | 2,451 | — | — | — | — | — | 3,137 |
| R66Resize224+MorphologicalDilation | 4,496 | — | — | — | 360 | 1,997 | 5,049 | — |
| R67Resize224+MorphologicalErosion | 4,429 | — | — | — | 363 | 2,150 | 5,073 | — |
| R68RandomScale+RandomCrop224 | 3,663 | 1,140 | 3,015 | 2,612 | 1,186 | 1,123 | 4,966 | 2,727 |
| R69Resize224+RandomSizedCrop | 4,313 | 1,872 | 3,295 | 2,020 | 1,610 | 2,202 | 5,022 | 3,512 |
| R70Resize224+CoarseDropout | 4,362 | 2,658 | 3,505 | 1,897 | 1,905 | 2,201 | 5,031 | 3,918 |
| R71Resize224+Emboss | 4,159 | 2,028 | — | — | 1,652 | 2,391 | 5,042 | 3,688 |
| R72Resize224+PixelSpread | 3,524 | — | — | — | — | — | 5,083 | — |
| R73Resize224+ShiftScaleRotate | 3,842 | 1,947 | 2,544 | 1,352 | 1,412 | 2,256 | 5,067 | 3,581 |
Measurement setup and limits
The timeline illustrates input stalls with a fixed model step duration.
g2-standard-16, nvidia-l4; 10,000 selected ImageNet JPEGs. Batch size 256, 15 workers, prefetch factor 2; persistent workers enabled. Output: cuda float16, BCHW 256×3×224×224.
Seeds: 137, 138, 139. Each observation follows 1 warm-up batch and times 32 batches, ending with CUDA synchronization. Pipeline construction and worker startup are outside throughput timing; prefetch effects remain. JPEG files are prewarmed, so this measures filesystem reads and decoding with a warm page cache.
NVML samples peak process GPU memory every 50 ms, from pipeline construction through final synchronization and cleanup. Brief peaks can be missed. The measurements include no model and do not establish training speed or augmentation quality. Seeds do not guarantee identical augmentation draws across libraries. Observed ranges describe variation between runs; they are not confidence intervals.
Run 03b5fd6c6fab74fd1029fc88012e03cb4b51d0a2e3c4b8be59ead6928efae6ee · 1323 measurements · measured source 3859d93 · machine-readable results · paper and methodology. This selected run differs from the paper's frozen results.

