dartraiden / NVIDIA-patcher

Adds 3D acceleration support for P106-090 / P106-100 / P104-100 / P104-101 / P102-100 / CMP 30HX / CMP 40HX / CMP 50HX mining cards.
194 stars 25 forks source link

Questions about CMP 90 HX #45

Open arsevka opened 8 months ago

arsevka commented 8 months ago

I want to buy cmp 90hx. Any information on whether nvenc/screen recording via geforce experience works on it? Does this graphics card have power limiting issues?

Skylord4321 commented 8 months ago

well i did own( and using it right now), from what i can tell it didn't because it could be a problem with the lastest driver image

Skylord4321 commented 8 months ago

and this card does have power limit lower than a normal 3080/3070 ti

Skylord4321 commented 8 months ago

huh interesting, however, there is another question i want to ask.

The driver version 537.58 you just added recently, what is that for?

On Wed, 11 Oct 2023 at 01:55, Alexander Gavrilov @.***> wrote:

Nothing changed in the driver except the signature, so the reason is something else

— Reply to this email directly, view it on GitHub https://github.com/dartraiden/NVIDIA-patcher/issues/45#issuecomment-1756478664, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOWGO2CIEBDJVDXIVA3QGYTX6XOABAVCNFSM6AAAAAA5YRX4USVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTONJWGQ3TQNRWGQ . You are receiving this because you commented.Message ID: @.***>

Skylord4321 commented 8 months ago

Interesting, mind if I ask can I try that driver? And from what I can tell, you can't change the power limit, right?

dartraiden commented 8 months ago

The driver version 537.58 you just added recently, what is that for?

Just an actual version.

it could be a problem with the lastest driver

Try the latest version.

you can't change the power limit

Currently I don't know how to fix this.

arsevka commented 8 months ago

Thanks for answers, I bought Asus CMP 40HX for testing. Judging by the number of blocks, this is a 2060S, the latest driver works fine, nvenc works, in-game screen recording in Geforce Expierence works. No power limit issues DLSS can be activated, there is a deterioration in the picture, but the fps does not change. RTX can be enabled in games, but the picture does not change, but FPS is greatly reduced. CB2077 1080p ultra settings no rtx/dlss, fsr quality ~50-60 fps

I have a question about Shield streaming, there is no “Shield” tab in GFe. Is it possible to activate it to use Moonlight?

Also I`m using Stable Diffusion, and there's a problem, the performance of the 40hx is 2.36 times worse than the p106. P106 creates 512x512 images in 30 seconds, 40HX does the same in 1m 11s. SD requires cuda, so it works on the original p106/40hx driver, and problem is not in the patch, I tried the original latest and old drivers 461.92 and got same results. Of course, the real 2060S is several times faster than the 1060, but why 40hx getting so bad results? Maybe it tries to use rt cores? Maybe theres a way to disable dlss/rt featues of card or spoof it like GTX one? Upd: fixed Stable Diffusion perfomance issue by adding --no-half to COMMANDLINE_ARGS

40hx

Skylord4321 commented 8 months ago

weird to ask but how did your gpu-z say that you have PCIe x16 1.1 @ x16 1.1 instead of PCIe x16 4.0 @ x16 1.1?

arsevka commented 8 months ago

I soldered missing capacitors on pcb, 1.1 is actually maximum version for this gpu

dartraiden commented 8 months ago

PCIe x16 4.0 @ x16 1.1

The text before the @ is what the video card supports (on-chip support, soldered pci-e bus elements, etc). The text after @ is the current operating mode (combination of what both the video card and motherboard supports).

Mining cards supports only PCI-E 1.1, so without soldering capacitors there will be (implying that the video card is plugged into a slot with 16 PCI-E lines) x4 1.1 @ x4 1.1. After soldering the capacitors there will be x16 1.1 @ x16 1.1. Even if the motherboard supports PCI-E 4.0, the resulting mode will be 1.1 because, as already said, it is a combination of what both the video card and motherboard supports.

dartraiden commented 8 months ago

I have a question about Shield streaming, there is no “Shield” tab in GFe. Is it possible to activate it to use Moonlight?

GFE recognizes the card based on DeviD, so does not allow these features to be used.

Maybe theres a way to disable dlss/rt featues of card or spoof it like GTX one?

I'll leave this ticket open, maybe someone will find a solution

arsevka commented 8 months ago

Mining cards supports only PCI-E 1.1, so without soldering capacitors there will be (implying that the video card is plugged into a slot with 16 PCI-E lines) x4 1.1 @ x4 1.1.

Should be x16 1.1 @ x4 1.1 this means gpu supports 16 lines of pcie. For example P106-090 / P104 / P102 can use only x4 lanes, and soldering caps won`t change anything, just search for gpu-z screenshots.

GFE recognizes the card based on DeviD, so does not allow these features to be used.

There is no way to trick it like it was done with shadowplay? Or maybe app patch?

operador777 commented 8 months ago

Shield streaming

Shadowplay is easy to bypass because it is an independent feature of GeForce Experience that can be activated using launch commands. The Nvidia Shield and other GFE features appear to be integrations and won't activate unless GFE detects a compatible graphics card.

In my free time, I am researching and attempting to create a modification for GeForce Experience to make it compatible with CMP graphics cards and other graphics cards that are incompatible.

Skylord4321 commented 8 months ago

Mining cards supports only PCI-E 1.1, so without soldering capacitors there will be (implying that the video card is plugged into a slot with 16 PCI-E lines) x4 1.1 @ x4 1.1.

Should be x16 1.1 @ x4 1.1 this means gpu supports 16 lines of pcie. For example P106-090 / P104 / P102 can use only x4 lanes, and soldering caps won`t change anything, just search for gpu-z screenshots.

GFE recognizes the card based on DeviD, so does not allow these features to be used.

There is no way to trick it like it was done with shadowplay? Or maybe app patch?

so what about flashing the bios chip? i can try it out and see if it does change anything

operador777 commented 8 months ago

Mining cards supports only PCI-E 1.1, so without soldering capacitors there will be (implying that the video card is plugged into a slot with 16 PCI-E lines) x4 1.1 @ x4 1.1.

Should be x16 1.1 @ x4 1.1 this means gpu supports 16 lines of pcie. For example P106-090 / P104 / P102 can use only x4 lanes, and soldering caps won`t change anything, just search for gpu-z screenshots.

GFE recognizes the card based on DeviD, so does not allow these features to be used.

There is no way to trick it like it was done with shadowplay? Or maybe app patch?

so what about flashing the bios chip? i can try it out and see if it does change anything

You can try to change the BIOS; I attempted it a couple of times through software, but it was impossible for me because when flashing the software, it checks the board ID and PCBID of the graphics card, so it aborts the operation. The recommended approach would be to flash it with an external programmer like CH341 and experiment with different BIOS versions.

You might want to consider trying the BIOS versions from the ASUS RTX 2060 Super up to the 2070 Super (note that this is for the CMP 40HX). If you do this, please keep us informed of your progress.

Skylord4321 commented 8 months ago

n try to change the BIOS; I attempted it a couple of times through software, but it was impossible for me because when flashing the software, it checks the board ID and PCBID of the graphics card, so it aborts the operation. The recommended approach would be to flash it with an external programmer like CH341 and experiment with different BIOS versions.

You might want to consider trying the BIOS versions from the ASUS RTX 2060 Super up to the 2070 Super (note that this is for the CMP 40HX). If you do this, please keep us informed of your progress.

thank you for saying, yes I do have an SPI programmer but not the cmp 40hx, I got cmp 90hx instead, I will gladly inform the progress

Skylord4321 commented 8 months ago

I'm back, the cmp90hx using IS25WP016, Maximum size is 16Mbit (2 MB): https://datasheetspdf.com/pdf/1036627/ISSI/IS25WP016/1

image

I have tried to flash it with different bioses and still continue doing it. Here is the list: image

These are marked as failed because they do not appear in the Device Manager. I will come back with more results, so any recommended Bios would be appreciated.

Oh one more thing i want to ask, which software that allows you to modify the power (via vbios) so that i can test it out as well

arsevka commented 8 months ago

What you want to achieve by flashing firmware from normal card? Video outputs and pci lanes version locked inside silicon, doesn't it? Doesn't gpu checks bios signature?

If it were possible to unlock the capabilities of a video card using BIOS firmware modification, our Chinese friends would sell these video cards, like they doing with mobile gpus, laptop cpu motherboards, laptop cpus with 1151 adapter and etc

Does your 90hx has soldered caps for 16 lines? What about powerlimit? Any benchmarks / in-game benches? how it performs? Have you tried to use stable diffusion on it? GPU-Z screenshot please.

Skylord4321 commented 8 months ago

oh right the bios signature, that's completely slipped my mind, thanks for reminding me, I was flashing it so that I can check if I can change the TDP of it. here is the GPU-Z image image

I gave it to the repair man this morning to apply the missing capacitors.

arsevka commented 8 months ago

It would be interesting to see your video card with 16 lines at work, have you tried running at least a furmark with 4 lines? How much FPS does it give?

Skylord4321 commented 8 months ago

I did run some tests but i forgot to save the screenshot, unfortunately, it runs worse than my 5700g apu

arsevka commented 8 months ago

Yea, that's bad, I also got 70HX today, soldered capacitors, and got very bad performance ~3 fps in furmark and 15 fps in Stray, you can see my new issue I opened about it. In my case this is not power limit or something, because frequencies are high, looks like Ampere CMP cards have limited in 3d loads. Btw 40HX performs same as 2060S, no problems, just works as it should.

Skylord4321 commented 8 months ago

Yea, that's bad, I also got 70HX today, soldered capacitors, and got very bad performance ~3 fps in furmark and 15 fps in Stray, you can see my new issue I opened about it. In my case this is not power limit or something, because frequencies are high, looks like Ampere CMP cards have limited in 3d loads. Btw 40HX performs same as 2060S, no problems, just works as it should.

so you can't control the frequencies via afterburner?

i did also noticed your cmp 70hx power draw image

arsevka commented 8 months ago

so you can't control the frequencies via afterburner?

I can, but gpu is limiting its performance by something.

The power consumption is most likely due to the fact that the video card is always running at boost frequency, and never tries to idle.

And in my case it’s not idling, most likely due to 100% video engine load as GPU-Z reports.

arsevka commented 8 months ago

SD requires cuda, so it works on the original p106/40hx driver, and problem is not in the patch, I tried the original latest and old drivers 461.92 and got same results. Of course, the real 2060S is several times faster than the 1060, but why 40hx getting so bad results? Maybe it tries to use rt cores? Maybe theres a way to disable dlss/rt featues of card or spoof it like GTX one?

Fixed it myself, 40HX cuda half-precision was cut off, and only single-precision can be used. Adding --no-half to COMMANDLINE_ARGS fixes bad perfomance

Skylord4321 commented 8 months ago

ok I have come back with the cmp ,i did test it with Furmark and i got 9 fps average, unfortunately, i couldn't test the pcie lane because i don't have a motherboard that has pcie x16 lane so i have to wait.

As for your cmp 70hx did you try that command line?

arsevka commented 8 months ago

No, this doesn't work on 70hx, I didn't get it to work, nothing works, 3d, neural networks and stable diffusion have very bad performance, like it's gt710 or gt210. So I decided to sell it.

9 fps in furmark, yea, but 40hx have 100 fps with just single x1 line

Skylord4321 commented 8 months ago

Ok that's good to know, I will wait for my new motherboard, meanwhile, I will look around for anything about these 70/90 problems.

Skylord4321 commented 8 months ago

I have found information regarding these problems : CMP 50HX 90HX避雷贴,为什么不要接盘 图是偷的,不碍事。鉴于一些矿卡JS乱吹50HX、90HX等CMP卡,故特地开一贴详细说明。 为什么不要接? 1、阉割半精:P106吧魔改驱动的无悔与流年大佬已经测过了,CMP系列全系列阉割了半精和Tensor Core,基本和炼丹无缘,光追也被阉割。 2、阉割单精:30HX和40HX单精还算健全,故可以用来打游戏。而50HX和90HX只保留了挖ETH需要的双精,单精也被阉割了,性能极差!不要幻想这些卡补过电容到X16能达到游戏卡水平,更不要相信某些JS说能持平同核心游戏卡2080TI、3080! 50HX和90HX这种单精半精全残废的玩意就是只能挖K的废铁!跑分都是很低的,没有JS敢贴出跑分来……别看他现在卖一个卖七八百,一个卖上千,事实上没有K潮,这俩玩意现在的价格不抹零都是没任何必要购买。 写于23.5.24 关于P10X系列K卡问题,有人问了,就补充一下 1、打游戏:只能考虑P106/30HX/40HX。原因就是它们有完整的1.1 X16通道。目前性价比最高P106白菜价,30HX比P106强的不多贵的不少。P104和P102不能补电容到X16通道,因为VBios限死了X4通道,而CMP没有限制,所以30HX和40HX可以补到X16。 2、炼丹:P104和P102的半精是健全没阉割的,不爆显存的前提下可以用于轻度炼丹,P104目前100来块还算便宜,搭配秋叶佬的包可以让很多人玩出图。当然30HX和40HX用单精也不是不能搞。 3、P104和P102不能打游戏,为什么跑分这么高?因为跑分本质是加载进显存后渲染,受pcie带宽影响很小,跟打游戏不一样……正常游戏卡少有阉割通道的,所以跑分可以基本代表游戏性能。这玩意就完全不同了,事实上p106也有微量损耗,40hx损耗也有小部分,只是它们是x16通道,损耗相对可以接受罢了。 修改补充于23.6.17 关于看视频:所有P10X和CMP系列K卡除了40HX以外都阉割了视频编解码器,别想看电影了!40HX是一个奇迹,保留了视频编解码器,相关性能和Turing同代显卡应该持平。 视频解码补充于23.7.18

This is the information I found from xianyu (basically Chinese Facebook marketplace)

Skylord4321 commented 8 months ago

I'm sorry if they are using weird "phrases" when using a translator because of translation error or they hate Nvidia

Skylord4321 commented 8 months ago

From what I guessing, the 50/70/90 hx has single and half float points fused off

tudoanh commented 7 months ago

From what I guessing, the 50/70/90 hx has single and half float points fused off

Really? So basically we can not run Stable Diffusion on those cards? So sad :(

Skylord4321 commented 7 months ago

That is just my guessing, right now I'm trying to find a way to prove if the floatpoint fused off was true

shpinog commented 6 months ago

Вы можете попробовать ваши карты в linux, возможно там не будет ограничений по 3D нагрузке. Так же попробуйте это там с версией драйвера 470.161.03, 495.46, 515.76.

avmarshall commented 6 months ago

Hi my CMP 70 HX Problem drivers mining 0’Pyrin and Kalrsen fix problem please! Cuda core not found

twobombs commented 5 months ago

Owner of two CMP50HX cards. I run LLMs on linux with normal drivers and cyberpunk on windows in fullhd in the 60s with modified NVidia drivers and FSR3 mod. https://github.com/dartraiden/NVIDIA-patcher

I also created a list with INT/FP performance for the tesla/mining cards that I owned at that point in time, more were added since, but this list should get the point across; this card is a steal at $160 a piece as a research device that requires a lot of INT and some memory. was looking here because I'm looking into flashing the CMP50HX to use the 20-series bridge. yeah, that's a long shot there

Screenshot from 2024-01-24 09-06-53 source: https://cuda-z.sourceforge.net/

GermanAizek commented 3 months ago

@dartraiden, I decided to ask why people are flashing CMP 90HX bios from RTX 3080? What is the advantage?

References: https://vlab.su/viewtopic.php?f=48&t=139106&p=1037859 (good revert attempt) https://internetbaraholka.ru/novosti-ot-olega/949-proshivka-videokarty-cmp-90hx.html (bad attempt reflash)

bah86 commented 3 months ago

@dartraiden, I decided to ask why people are flashing CMP 90HX bios from RTX 3080? What is the advantage?

References:

https://vlab.su/viewtopic.php?f=48&t=139106&p=1037859 (good revert attempt)

https://internetbaraholka.ru/novosti-ot-olega/949-proshivka-videokarty-cmp-90hx.html (bad attempt reflash)

Id чипа разные. После прошивки код 43. Проверено. Редактировать BIOS не получится, защита считывает контрольную сумму и если она не совпадает, карта не запуститься.

aleksandraleksandrov666666 commented 1 month ago

Кому-то удалось завести cmp 50 hx?