AI Benchmark

Ignatov Andrey

Mayelana nalolu hlelo lokusebenza

Isizukulwane Sesithombe Se-Neural, Ukubonwa Kobuso, Ukuhlelwa Kwesithombe, Ukuphendulwa Kwemibuzo...

Ingabe i-smartphone yakho iyakwazi ukusebenzisa ama-Deep Neural Networks akamuva ukwenza le misebenzi neminye eminingi esekelwe ku-AI? Ingabe inayo i-AI Chip ezinikele? Ingabe ishesha ngokwanele? Qalisa i-AI Benchmark ukuze uhlole ngobungcweti Ukusebenza kwayo kwe-AI!

Izinga lefoni lamanje: http://ai-benchmark.com/ranking

I-AI Benchmark ikala isivinini, ukunemba, ukusetshenziswa kwamandla kanye nezidingo zenkumbulo zokhiye abambalwa be-AI, i-Computer Vision kanye namamodeli e-NLP. Phakathi kwezixazululo ezihloliwe kukhona Ukuhlelwa Kwezithombe kanye nezindlela Zokuqashelwa Kobuso, amamodeli e-AI enza isithombe se-neural nokukhiqizwa kombhalo, amanethiwekhi e-neural asetshenziselwa i-Image/Video Super-Resolution kanye Nokuthuthukisa Isithombe, kanye nezixazululo ze-AI ezisetshenziswa ezinhlelweni zokushayela ezizimele nama-smartphones ngokoqobo- isikhathi Ukulinganisa Ukujula kanye Semantic Image Segmentation. Ukubonakala kwemiphumela yama-algorithms kuvumela ukuhlola imiphumela yawo ngemidwebo kanye nokwazi isimo samanje sezobuciko emikhakheni ehlukahlukene ye-AI.

Sekukonke, i-AI Benchmark iqukethe izivivinyo ezingama-83 nezigaba ezingama-30 ezibalwe ngezansi:

Isigaba 1. Ukuhlelwa, i-MobileNet-V3
Isigaba 2. Ukuhlelwa, Ukuqaliswa-V3
Isigaba 3. Ukubonwa kobuso, i-Swin Transformer
Isigaba 4. Ukuhlelwa, I-EfficientNet-B4
Isigaba 5. Ukuhlelwa, i-MobileViT-V2
Izigaba 6/7. Ukwenziwa Kwemodeli Ehambisanayo, 8 x Ukuqaliswa-V3
Isigaba 8. Ukulandelelwa Kwezinto, YOLO-V8
Isigaba 9. I-Optical Character Recognition, i-ViT Transformer
Isigaba 10. Semantic Segmentation, DeepLabV3+
Isigaba 11. I-Parallel Segmentation, 2 x DeepLabV3+
Isigaba 12. Isegimenti Semantic, Ingxenye Noma yini
Isigaba 13. Ukufiphaliswa Kwezithombe, IMDN
Isigaba 14. Image Super-Resolution, ESRGAN
Isigaba 15. Image Super-Resolution, SRGAN
Isigaba 16. Ukuhlehla Kwezithombe, i-U-Net
Isigaba 17. Ukulinganisa Ukujula, i-MV3-Ukujula
Isigaba 18. Ukulinganisa Ukujula, MiDaS 3.1
Isigaba 19/20. Ukuthuthukiswa Kwesithombe, DPED
Isigaba 21. I-ISP yekhamera efundiwe, i-MicroISP
Isigaba 22. I-Bokeh Effect Rendering, i-PyNET-V2 Mobile
Isigaba 23. I-FullHD Video Super-Resolution, XLSR
Isigaba 24/25. I-4K Video Super-Resolution, VideoSR
Isigaba 26. Ukuphendulwa Kwemibuzo, MobileBERT
Isigaba 27. I-Neural Text Generation, Llama2
Isigaba 28. I-Neural Text Generation, GPT2
Isigaba 29. I-Neural Image Generation, Ukusabalalisa Okuzinzile V1.5
Isigaba 30. Imikhawulo Yenkumbulo, ResNet

Ngaphandle kwalokho, umuntu angalayisha futhi ahlole amamodeli abo okufunda okujulile e-TensorFlow Lite Kumodi ye-PRO.

Incazelo enemininingwane yokuhlolwa ingatholakala lapha: http://ai-benchmark.com/tests.html

Qaphela: Ukusheshiswa kwezingxenyekazi zekhompuyutha kusekelwa kuwo wonke ama-SoC eselula anama-NPU azinikele nezisheshisi ze-AI, okuhlanganisa i-Qualcomm Snapdragon, MediaTek Dimensity / Helio, Google Tensor, HiSilicon Kirin, Samsung Exynos, kanye nama-chipsets e-UNISOC Tiger. Kusukela ku-AI Benchmark v4, umuntu angakwazi futhi ukunika amandla ukusheshisa kwe-AI okusekelwe ku-GPU kumadivayisi amadala kuzilungiselelo ("Sheshisa" -> "Vumela Ukusheshisa kwe-GPU" / "Arm NN", i-OpenGL ES-3.0+ iyadingeka).

Yini entsha

1. LiteRT (TFLite) runtime updated to version 2.18.
2. Updated Qualcomm QNN, MediaTek Neuron, Samsung ENN, TFLite NNAPI, GPU and Hexagon NN delegates.
3. Bug fixes and performance improvements.