NVIDIA GeForce RTX 4090

Ada, AD102, 16384 shaders, PCIe x16 4.0

Thanks to @MysteryDove

Benchmark

vsmlrt v12

• processor clock @ 2860 MHz
• memory clock @ 1406.25 MHz
• driver 527.56
• Windows 11 21H2
• VapourSynth R57

FP16

1920x1080 rgbs

Measurements: FPS / Device Memory (MB)

model	ORT_CUDA 1 stream	TRT 1 stream	TRT 2 streams
dpir gray	9.093 / 1845	23.95 / 975	24.527 / 1297
dpir color	8.688 / 1749	21.92 / 1073	24.212 / 1413
waifu2x upconv7	20.15 / 5905	39.035 / 2354	53.406 / 4029
waifu2x upresnet10	13.445 / 2814	29.411 / 2147	37.001 / 3439
waifu2x cunet	6.981 / 8532
cugan	6.869 / 8719	20.511 / 5490	24.018 / 9579
realesrgan	14.188 / 2346	28.518 / 2080	35.771 / 2970
rife (1920x1088, model=44)	95.52 / 1609	138.513 / 1319	208.345 / 1653

vsmlrt v14.test2

• driver 545.84
• Windows 10 22H2
• VapourSynth-classic R57.A8

Waifu2x.swin_unet_art

1920x1080 rgbs

PSNR is tested on a private set of samples compared to FP32.

Measurements: FPS / Device Memory (MB)

precision	TRT 1 stream	TRT 2 streams	PSNR
fp16	4.91 / 7399.8	5.40 / 14351	65.3
bf16	4.70 / 7797.1	5.10 / 15142	53.7