[显卡推荐]madVR需要什么级别的显卡?

Toxin-avatar

Toxin

2021-08-14T13:32:46+00:00

有用过madVR的朋友们:
目前用16系笔记本外接,进入夏天后风扇起飞听着很心烦。
那么什么样的显卡可以在暂不考虑hdr映射、仅考虑sdr缩放1080p->4K或720p->4K下,画面倍线里nguAA(upscaling refinement里勾选cripsen edges及thin edges)以及nguSharp开启very high时,渲染frame time在30ms以下呢?

2070s/3060ti/6600xt级别?
还是3070/6700xt级别?

谢谢各位!
陳彥樺-avatar

陳彥樺

4K madvr拉满+补帧,3080占用率大概在60%-70%,供你参考
Nightcore(Алексей)-avatar

Nightcore(Алексей)

拉满的话好多显卡都废了
Maid sama-avatar

Maid sama

3090
6900XT
好像還不夠
SC00ba-avatar

SC00ba

这我熟
按你的条件 首先必须得是桌面端显卡
然后3080起步
由于作者不怎么对a卡进行优化
所以无脑选n卡。
以上只是4k 60刷新的视频 视频帧率再高 3090也不够
samgu_lol-avatar

samgu_lol

4k 拉满ngu very high能拉爆,99占用,风扇起飞,我是evga 3080ti
Blue Fire God-avatar

Blue Fire God

6700xt含以上
因为我5700xt开2k upscaling sharpness差不多能拉满
而5700xt也就比6600xt弱一丢丢
你开4k那肯定是6700xt起步[s:ac:哭笑]
Toxin-avatar

Toxin

谢谢各位的回复!不过我还有一些不确定的地方希望能有更详细的建议。
我平时主要看的电影和视频是24帧和30帧, 目前的1660可以做到37ms在nguSharpVH,满足一部分24fps内容的播放,就是有点吵,nguAAvh则需要平均48ms才能处理一帧 ,不考虑补帧,这也是我希望单帧的渲染时间尽量能控制在30ms以下的原因。
在[url=https://www.chiphell.com/thread-2274087-1-1.html]chiphell一位用户的帖子[/url]里,3080仅作倍线时的单帧渲染时间在16ms到20ms之间,这对没有看60fps视频需求的场景来说无疑是一种巨大的性能浪费,因此才提出询问2070s~3070区间的性能表现。所以可以的话,希望各位能具体秀一下1k视频倍线到4k时的frametime,也方便日后做一个参考。

反正台机不用放在桌面上,风扇起飞了就扔墙角去

另一个令我在意的地方在于对这些内容谷歌的时候得到的一些难以验证的内容:

“madvr使用16bit integer纹理,所以比起对游戏性能提升很大的fp32,显卡的int16运算能力在这里更重要;”

如果上述推断是真的,那么A卡和RTX20系可能要比同规格的GTX10系或RTX30系更适合madvr,这些卡的int16运算能力要更强一些。
来自5ch,有一个madvr相关的专楼 ...
例:
[quote]ゲームやらずにmadVR用でグラボ使っているぞ madVRはINT16で処理してるから実質Radeon一択になったな RTX2000のINT16はFP32の2倍の処理能力があったけど RTX3000でINT16はFP32の等倍に戻ったしな RadeonはVEGA世代からずっとINT16はFP32の2倍ののままで RDNA2でもそれは変わらない[/quote]
不过类似的说法我只在5ch上看到,反而是英语论坛用“madvr”,“int16”,“fp32”,“rdna”等关键词搜索并没有什么收获,只不过是确认了madvr确实有默认用16位整数处理画面的流程:
madvr使用int16纹理 ...
摘自[url]http://forum.doom9.net/showthread.php?p=1632165[/url]:

I'm using 16bit integer textures by default, which have the benefit of using all their accuracy on the important 0..1 range. Yes, 16bit floats are relatively fast, too, but they stretch their accuracy over a *much* wider range, which means that in the area which counts the most (16-235) they have a much lower accuracy compared to 16bit integer. If I had to use floating point textures, I would have to go 32bit floating point by default, in order to not compromise accuracy. So the decision for me is not 16bit integer vs 16bit float, but it is 16bit integer vs 32bit float, and that's a quite big difference in performance.

Ok, if I went with floating point, people could compromise by using 16bit float instead of 32bit (trade quality for performance). They'd probably get identical performance to the current madVR logic, but with slightly faster subtitle rendering. On the negative side, they'd get less accuracy at the same speed as the current madVR logic. I think that's a bad compromise. Furthermore, if you look at the high quality option, I'd have to use float32 instead of int16, which costs twice the memory bandwidth. So the memory bandwidth cost of doing Jinc3 AR for a "no compromise" setup would double. All of that just to get a small boost in subtitle rendering performance? Sounds like a bad idea to me...
以及确认了Madshi这几年研究出来的madVR Envy播放器里面塞的是2080Ti,豪华版里是3080 看一眼信用卡余额

如果有持有6700xt及以上级别A卡的朋友愿意分享一下使用情况的话,将是极大的帮助。
dontpickme-avatar

dontpickme

不补帧还好,不过我这3080ti开了以后都不好拉进度条,一拉就啸叫一秒,打游戏都没这声音。
Toxin-avatar

Toxin

[url=http://forum.doom9.net/showthread.php?t=176013]doom9上有一个持续追踪N\A\I家新驱动对madvr影响的帖子[/url],
里面提到Amd的rdna显卡存在去交错和黑阶的bug ...
Known limitations of Radeon RX 5000 and 6000 GPUs (older generations are fine):
• Hardware deinterlacing is broken in every driver version -> It is poor quality when applied to
interlaced video and it is not adaptive anymore, which is a NO-GO for HTPC usage.
You can check this issue yourself: Download this short MBAFF sample file
The Hardware Deinterlacer should be aware that the characters above "WWW.D88.COM" are
progressive and should not be deinterlaced, otherwise you will see flickering like in this picture
See here for more details
• Color levels (full 0-255 vs. limited 16-235) are broken. That means limited 16-235 which is the
standard range for videos are interpreted as 0-255 and this leads to elevated blacks and dimm
whites.
This issue can be best seen on displays which want full range 0-255, like TVs in PC Mode (PC
mode for full 4:4:4/RGB support, so you do not use color subsampling like 4:2:2 or 4:2:0)
Also overriding the range for test purposes in MPC-BE (video tab -> Output range) does not
take effect.
Dear AMD, if any of your driver programmers read this: Please fix this fundamental video
bugs .These are there since RX 5000 series launch.

看来rtx 和vega 是使用madvr的最佳选项了
Ugly_J3st3r-avatar

Ugly_J3st3r

madvr拉满根本没有必要
MonoTyp3-avatar

MonoTyp3

有没有madvr设置一键导入的那种,以前用过一个换机子找不到了,看着一堆英文设置脑阔痛[img]http://img.nga.178.com/attachments/mon_201209/14/-47218_5052bc521c04b.png[/img]
Toxin-avatar

Toxin

[quote][pid=542815300,28108660,1]Reply[/pid] Post by [uid=60979474]STHGINKRA[/uid] (2021-08-19 19:30):

madvr拉满根本没有必要[/quote]这个倒确实是大实话,实际上luma升频算法里ngu medium的效果就已经非常好了,继续拉高则可以减轻拉伸带来的振铃等效果——但那可能需要靠近屏幕仔细观察才能看到,起码在1.5m视距上看55寸屏很难有明显的感知。
Toxin-avatar

Toxin

[quote][pid=542827964,28108660,1]Reply[/pid] Post by [uid=33709225]特拉迪格斯[/uid] (2021-08-19 20:23):

有没有madvr设置一键导入的那种,以前用过一个换机子找不到了,看着一堆英文设置脑阔痛[img]http://img.nga.178.com/attachments/mon_201209/14/-47218_5052bc521c04b.png[/img][/quote]一键导入没找到,不过其实只要复制粘贴根目录下几个文件名里带‘setting’的文件应该就可以了吧
Moondog--avatar

Moondog-

不拉满2080也够用,拉满3090也不够用,亲测
Brates-avatar

Brates

我的1060还在用

另外单纯1080 upscale 4k的话
github上有个anime4k的项目做的很好建议体验下
Toxin-avatar

Toxin

Reply to [pid=542914700,28108660,1]Reply[/pid] Post by [uid=58568]摸鱼的[/uid] (2021-08-20 07:33)
mpv确实是很不错的选项,就是界面不太友好配置有点麻烦,不过很省钱包,毕竟拉升缩放上只需要mpc+madvr+nguHigh不到一半的效能就可以有类似的效果,很不错。
Anime4K确实很”强力“,但是过于强力了,即便是力度最小的faithful滤镜组对画面细节的改动也很强,而且也有人认为[url=https://redvice.org/2019/on-anime4k/]”Anime4K根本就不是升频算法“[/url],相比之下我觉得fsrcnnx更好一些。

不过我最近在找一些老动画片的片源,其中很多都是些丑陋的480idvdraw,这个时候Anime4K强力处理后的画面感官上要比其他单纯升频的滤镜好一些,有时会不如madvr里ngu+rca。

[quote][pid=542911976,28108660,1]Reply[/pid] Post by [uid=60131852]benzhou222[/uid] (2021-08-20 07:01):

不拉满2080也够用,拉满3090也不够用,亲测[/quote]其实我一直在考虑amd,毕竟同档次都要便宜一些,不过看doom9上的讨论和样张,rdna在驱动上似乎有点问题。
24p内容纯拉升缩放似乎2070s级别就可以了,高帧率HDR内容或加入rca或者用double again的话那确实顶配也不够用。
Brates-avatar

Brates

我用anime4k也看了一些老片子 效果都还可以

好多时候还是暗场的处理 banding ringing blocking combing之类的画面瑕疵比较影响观感