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.

Ready batches per second and peak process GPU memory on 35 common recipes. Each metric takes the median across seeds, then across recipes. The dashed line marks AX's batch rate.
Ready batches per second and peak process GPU memory on 35 common recipes. Each metric takes the median across seeds, then across recipes. The dashed line marks AX's batch rate. Scroll horizontally to see the full chart. Open the image for full size.

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.

Benchmark metric

Higher throughput is better. Values are medians across seeds. Hover for the observed range. A dash means no measured result.

R01Resize2244,6912,6653,8543,4252,3292,3735,0564,176
R02RandomCrop2244,4214,3194,1204,4642,2562,2556,3874,197
R03RandomResizedCrop4,7062,7143,5872,3361,9231,9556,0653,904
R04HorizontalFlip3,9554,2874,1793,6452,1282,2036,4174,291
R05VerticalFlip4,2794,2554,2203,4952,0732,2076,4014,146
R06Pad+RandomCrop2244,2443,8913,5473,1471,7921,8484,9964,022
R07Rotate4,4094,0572,4991,2451,5202,2486,3794,192
R08Affine4,1062,6252,6521,2801,5332,2716,4333,902
R09Perspective4,1802,6502,6451,2621,5362,3066,1273,669
R10Elastic3,032—236227932,371——
R11ColorJitter3,569—9676899683313,511—
R12ChannelShuffle4,7514,1843,9862,8672,1132,1586,2514,192
R13Grayscale4,8044,1043,8413,1522,0212,1976,4164,100
R14RGBShift4,6553,833——2,0282,2596,1904,006
R15GaussianBlur4,3992,5212,1681,9041,2262,1346,3293,764
R16GaussianNoise3,376—1,2313,1321,7452,2136,372—
R17Invert4,9283,9404,2213,4832,0702,1826,3924,186
R18Posterize4,9683,9664,1473,0751,6622,2256,3714,029
R19Solarize4,6223,8203,5833,1681,9112,2126,4014,245
R20Sharpen4,481—2,3071,9411,2822,264——
R21AutoContrast4,1603,7242,7182,3441,8342,158—3,882
R22Equalize3,9763,7502,9731,2931,5142,2196,3883,872
R23Erasing4,726—3,8771,9321,7612,1236,293—
R24JpegCompression3,8683,5113,3282,3606682,0866,3403,611
R25RandomGamma4,5864,0543,0593,8211,7152,2476,4274,095
R26PlankianJitter4,340———1,9722,225——
R27MedianBlur3,758243——4057625,0803,916
R28MotionBlur4,213———1,4742,241——
R29CLAHE3,109———8122,1625,955—
R30Brightness4,6333,9543,8833,5622,0552,1516,3443,866
R31Contrast4,3803,6903,4282,6761,9652,1786,3002,439
R32Blur4,8513,341——1,0582,2116,3494,141
R33ChannelDropout4,7324,072——2,0592,1806,2894,264
R34LinearIllumination3,826———1,8162,174——
R35CornerIllumination3,991———1,5872,115——
R36GaussianIllumination3,789———1,5272,206——
R37Hue4,064—1,4038381,0892,1916,357—
R38PlasmaBrightness2,603———4812,127——
R39PlasmaContrast2,202———4722,154——
R40PlasmaShadow2,502———7972,225——
R41Rain3,979———1,5412,154——
R42SaltAndPepper4,750———1,5338006,332—
R43Saturation4,3543,798——1,9752,3636,3234,030
R44Snow3,733———1,0852,227——
R45OpticalDistortion3,817———1,3032,306——
R46Shear3,7562,6432,8391,5261,5542,2746,4363,822
R47ThinPlateSpline3,139———6432,306——
R48PhotoMetricDistort3,482—1,172635————
R49ColorJiggle3,640—1,2106879753333,461—
R50LongestMaxSize+RandomCrop2243,7291,2363,3802,4921,2631,2855,0232,950
R51SmallestMaxSize+RandomCrop2243,1878483,0792,5939809715,0252,403
R52Transpose4,9454,1584,0144,1681,6872,1306,3834,342
R53RandomRotate904,8464,2474,1493,8401,3892,2046,3834,368
R54RandomJigsaw4,442———1,8782,245——
R55EnhanceEdge4,4802,699—————3,995
R56EnhanceDetail4,4122,762—————4,046
R57UnsharpMask4,2522,276—————3,604
R58CenterCrop4,6584,2064,2405,9972,7442,6636,5064,358
R59Resize256+Crop4,5952,3423,6053,3022,0112,0925,0913,805
R60Resize224+CropAndPad4,2842,6732,9002,8352,2102,2645,0663,979
R61LetterBox4,6492,8763,7722,7622,3662,5935,0543,987
R62PadIfNeeded+RandomCrop2244,4044,1234,2754,0142,3282,1735,0074,400
R63Resize224+ChannelSwap4,5692,6573,5852,6362,1302,2665,0633,722
R64Resize224+Colorize4,2672,409——1,4172,2535,0583,531
R65Resize224+Dithering8952,451—————3,137
R66Resize224+MorphologicalDilation4,496———3601,9975,049—
R67Resize224+MorphologicalErosion4,429———3632,1505,073—
R68RandomScale+RandomCrop2243,6631,1403,0152,6121,1861,1234,9662,727
R69Resize224+RandomSizedCrop4,3131,8723,2952,0201,6102,2025,0223,512
R70Resize224+CoarseDropout4,3622,6583,5051,8971,9052,2015,0313,918
R71Resize224+Emboss4,1592,028——1,6522,3915,0423,688
R72Resize224+PixelSpread3,524—————5,083—
R73Resize224+ShiftScaleRotate3,8421,9472,5441,3521,4122,2565,0673,581

Measurement setup and limits

When a batch arrives late, the GPU sits idle and completes fewer steps in the same time.
When a batch arrives late, the GPU sits idle and completes fewer steps in the same time. Scroll horizontally to see the full chart. Open the image for full size.

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.