当前位置：网站首页>[Xingguang 04] 2022 deep learning GPU form

[Xingguang 04] 2022 deep learning GPU form

2022-07-18 03:10:00 【Yang Sier】

model memory Single sperm (FP32) Semi refined (FP16) explain
Tesla P4024GB11.76 T11.76 T Older Pascal framework GPU, about cuda11.x Previous algorithms that require large video memory are a very good choice
TITAN Xp12GB12.15 T12.15 T Older Pascal framework GPU, It is more suitable for getting started
1080 Ti11GB11.34 T11.34 T and TITANXp Contemporary cards , Also suitable for entry , however 11GB Occasionally, my video memory will be embarrassed
2080Ti11GB13.45 T53.8 T Turing Architecture GPU, Good performance , Among the older models, it is more suitable to calculate the mixing accuracy GPU. High cost performance
V10016/32GB15.7 T125 T The old generation of Professional Computing card king , High semi precision performance, suitable for mixing accuracy calculation
306012GB12.74 T about 24T If 1080Ti My video memory is just embarrassed ,3060 It's a good choice , Suitable for novice . Need to use cuda11.x
A400016GB19.17 T about 76T Both video memory and computing power are relatively balanced , Suitable for advanced process . Need to use cuda11.x
3080Ti12GB34.10 T about 70T Performance steel gun , If the demand for video memory is not high, it is a very suitable choice . Need to use cuda11.x
A500024GB27.77T about 117T Performance steel gun , If you feel 3080Ti Not enough video memory A5000 It's the right choice , And the semi actuarial force is high, which is suitable for mixing accuracy . Need to use cuda11.x
309024GB35.58 T about 71T Can be seen as 3080Ti Extended memory version of . Performance and video memory size are very adequate , Very applicable , Cost performance is the first choice . Need to use cuda11.x
A4048GB37.42 T149.7 T You can view it as 3090 Extended memory version of . Calculative sum 3090 flat , So choose according to the size of the video memory . Need to use cuda11.x
A100 SXM440/80GB19.5 T312 T A new generation of Professional Computing card king , There are no shortcomings except expensive . Video memory is large , It is very suitable for semi precise calculation , Because there is NVLink The blessing , The multi card parallel acceleration ratio is very high . Need to use cuda11.x

Reference resources ：AutoDL Help document /GPU The selection
Reference resources ：AutoDL Help document / Performance measurement

原网站

版权声明
本文为[Yang Sier]所创，转载请带上原文链接，感谢
https://yzsam.com/2022/199/202207151145593980.html

边栏推荐

猜你喜欢

随机推荐