Jump to content

[KEXT] Tired of low Geekbench scores? Use RadeonBoost!


CMMChris
 Share

133 posts in this topic

Recommended Posts

On 4/20/2020 at 6:15 AM, CMMChris said:

New version 1.2 added in first post. Vega10 support has been removed (Vega 56, Vega 64, Vega Frontier).

 

Reason:
People have observed fan control being broken when using this kext like it was the case prior to the release of macOS 10.14.5 Mojave. You remember? 10.14.5 was the macOS release that fixed the fan issues with Vega 56, 64 and Frontier cards. But you might also remember that exactly with this release the Geekbench scores have become lower as well.

 

Hey Chris thanks for this!

I have a vega FE but I put it on water so the fan control issue was never an issue for me, plus I'm on High Sierra on account of my {censored} old CAD software only runs properly on that. So can you flick me the kext version that still supported the FE?

Geekbench 5 is 66020 on openCL but might as well go for a wee bit more lol.

Link to comment
Share on other sites

On 4/23/2020 at 5:45 AM, Tiem said:

 

This results in a KP on the latest Cat betas (10.15.5). Looks like 10.15.5 might break all this stuff. The Navi chip is really just all over the place.

I had got KP at 10.15.5 beta2 too.

It had maximal 5 % increase by GeekBench 5 scores at 10.15.3/10.15.4, but no so constantly for my RX-580 in Asus Z87 hackintosh.

[Edit]

After add 0x67FF1002 in RadeonBoost.kext by OpenCore 0.5.6 booting:

1. No KP for FX-6300 hackintosh at 10.15.5 beta2

2. Significant increase of GeekBench 5 scores for RX-560 from 13~14K to 20~22K of OpenCL & Metal, respectively.

[Edit 2]

After repeated testing even without RadeonBoost my RX-560 scores still got around 24K for OpenCL and 26~27K for Metal. I don't know why !

Edited by jsl2000
Link to comment
Share on other sites

44 minutes ago, jsl2000 said:

I had got KP at 10.15.5 beta2 too.

It had maximal 5 % increase by GeekBench 5 scores at 10.15.3/10.15.4, but no so constantly for my RX-580 in Asus Z87 hackintosh.

Ya that’s zero gain. Run the test 10x and take the mean. Then take the rest with a grain of salt. I’ve had scores of 38k to 49k. There are big swings with GB. 

Link to comment
Share on other sites

Hi Chris, Great job! Thanks.

 

RadeonBoost.kext v1.0 in kexts/Other using Clover r5112 and PowerPlayTable from config.plist gave a big boost in Geekbench 5:

 

340017729_Screenshot2020-04-23at15_29_52.thumb.png.711d27ef64dfdb9438cac0f1c3b74489.png

 

No problems with the Vega 64 giving excessive fan noise either.

Link to comment
Share on other sites

With the kext and my RX 460 , 10.15.4,  i also get around 30% higher geekbench 5  OpenCL and Metal scores, So far so good.

 

1. I don't get any boost in other gpu compute benches like LUXMARK etc. and no boost of FPS in any Game with buildin FPS counter i testet.

2. I can get (Thanks to Chris!) same higher, 30%, Geekbench  5 scores - and only there - if i inject 3 device properties by Clover (Screenshoot)

So i get same boost also without the kext. AGPM (is within newer kext versions) has no effect on my RX 460 card.

Bildschirmfoto 2020-04-25 um 10.23.49.jpg

 

Chris tells that the reason for higher Geekbench scores (some get 50%++!) came by higher vram bandwidth/internal VRAM tranfers and access goes faster and Geekbench is sensitiv to that. I dont think so because Geekbench measures score by about 8 different small gpu compute tasks. Even if some are sensitiv to the vram bandwidth it tells not why other gpu compute tasks or benches show really zero effect - for me Luxmark speed excat same not even 1% difference. And Luxmark uses also big ammount of the VRAM  (up to 2 GB) of course.

So why most of us - if they have an boost more than 10% between 2-3 geekbench runs - cant measure any boost with other benches or apps which can be users as benches.

"FCP runs smoother" is for me or others not reproducible - we would need same source video and must do same things so compare.

I tested some games which shows same FPS with/without kext and Luxmark + other gpu compute benches - the 30% Geekbench boost wasnt shown in any, not even 1%.

 

I will NOT say the kext is placebo because Geekbench shows reproducible effect - but perhaps the boost comes by geekbench itself, geekbench optimises some on the fly in their gpu compute code if it found those kext injected parameters?

Really, if i get 30% boost with kext, others up to 100%!,  but cant find any boost in other  benches / apps something is to investigate.

Great would be others with same boost whoch find an bench / app whichs shows the booast by values and not "feels smoother, faster".

I used also Idigo Render Bench http://downloads.indigorenderer.com/dist/beta/IndigoBenchmark_v4.0.64.pkg, but also here no boost, same as without kext. HINT: Disable CPU as render device ;) Should bei gpu only like on sceenshoot.

 

Bildschirmfoto 2020-04-24 um 09-1.29.43.jpg

 

PS: Maybe oclBandwidthTest is the rigth tool to verify if the these "bandwidth/VRAM speed" may explain why geekbench boosts with the kext.

For me the bandwidth is even a little bit lower using the kext (i belive by AGPM content) by using device properties inject bandwidth is same again. With or without device properties (=boost kext)  i get same  80 GB/Sec in the oclBandwidthTest. 

Usage: Start Terminal, oclBandwidthTest --dtod --mode=shmoo 

Run the oclBandwidthTest at least 3 times (each run takes only about 2 Seconds ;) ), there will be differences between the runs, so you get an better average value so can compare with/ without kext.

(parameters gives an more intensitive internal bandwisth test , with out you get also PCI to / from gpu transferspeeds and an quick device to device(internal) bandwidth test.

 

 

 

 

 

 

oclBandwidthTest.zip

Link to comment
Share on other sites

@mitch_de interesting, i own a Sapphire Pulse RX 5600 XT and the only effective parameter for me is Force_Load_FalconSMUFW, PP_WorkloadPolicyMask was yet by default in 0x20.

And the only bench that shows me a improvement is Geekbench (bandwidth in Sobel test is doubled). Any other test gives me same results with or without kext/Injected device properties.

We will see, my thoughts are Apple is improving drivers, release by release so probably we don't need any big tweaking.

Edited by Xanny
  • Like 1
Link to comment
Share on other sites

Yep, without checking (against other apps than geekbench) it is unclear if the boost

1. is really happen  also in other apps 

Some investigation - i will look after the files (libs, amd drivers etc.) are loaded and used when geekbench is running compute - with boost kext and without. Perhaps thats the difference. Or, perhaps geekbench checks something / optimise something in their app which result in boost.

EDIT: Geekbench loads and uses same libs /AMD driver etc. 

 

@Xanny: Can you try to measure the bandwidth speed with boost (kext) and without. Perhaps , if you measure higher value for the bandwidth iit may be the reason. For me its not - RX 460 has same bandwidth, doesnt matter using kext or not.

Small oclBandwidthTest is in my last post (2 post above this). Your bandwidth will be much hogher than my around 80000. I think at least 120000. But the interest is only in the difference (perhaps) with kext (or device properities inject) and without.

 

 

Link to comment
Share on other sites

18 hours ago, mitch_de said:

@Xanny: Can you try to measure the bandwidth speed with boost (kext) and without. Perhaps , if you measure higher value for the bandwidth iit may be the reason. For me its not - RX 460 has same bandwidth, doesnt matter using kext or not.

Small oclBandwidthTest is in my last post (2 post above this). Your bandwidth will be much hogher than my around 80000. I think at least 120000. But the interest is only in the difference (perhaps) with kext (or device properities inject) and without.

 

 

I did the tests, there is an effective improvement of the bandwidth. i repeated the tests few times results are comparable, attached here my two tests you can check. As i stated before i just inject Force_Load_FalconSMUFW directly with clover config.plist, but the kext do the same. The SSDT someone is working on, inject many totally useless things for my Graphic card, so i do not tested it.

Anyway you should try to check if there is a difference between two tests one on windows, one on mac os, if there is a difference and in which sector bandwith or whatever, then find the right parameter for your card if any.

Bandwidthtest_with.rtf

Bandwidthtest_without.rtf

Edited by Xanny
typos
Link to comment
Share on other sites

Yeah doesn't work for some cards. Apple's SMU firmware for Navi appears to be incompatible with some of the Navi cards for some reason. Nothing you or I can do about that unfortunately. Apple has to fix up their {censored} and make my kext obsolete :)

Link to comment
Share on other sites

17 hours ago, Xanny said:

I did the tests, there is an effective improvement of the bandwidth. i repeated the tests few times results are comparable, attached here my two tests you can check. As i stated before i just inject Force_Load_FalconSMUFW directly with clover config.plist, but the kext do the same. The SSDT someone is working on, inject many totally useless things for my Graphic card, so i do not tested it.

Anyway you should try to check if there is a difference between two tests one on windows, one on mac os, if there is a difference and in which sector bandwith or whatever, then find the right parameter for your card if any.

Bandwidthtest_with.rtf

Bandwidthtest_without.rtf

 

Interesting results.

But it shows, that the maximum bandwidths are near the same for transfer sizes bytes more than around 20 KB. In your results the smaller transfersize speeds shows much , up to double , difference between with / without booster.

Bildschirmfoto 2020-04-28 um 08.09.08.jpg

 

 

I cross tested that on my RX 460 (booster gain 30%!) with an closer look to the smaller transfer sizes, because like you my maximum BAndwidths are also same (less than 2% diff with/ without booster) , nears same as yours. Between runs the smaller transfersizes vary much more than the bigger transfersite bandwidths.

For my RX 460 frist run was onyl about 10 MB/S

Device to Device Bandwidth, 0 Device(s)
   Transfer Size (Bytes)    Bandwidth(MB/s)
   1024                9.2
   2048                146.2
Second  run and all other runs around 49 MB/S

Device to Device Bandwidth, 0 Device(s)
   Transfer Size (Bytes)    Bandwidth(MB/s)
   1024                50.4
   2048                160.4

 

So, if you also repeated the test are you sure also Bandwidths of the smaller transfersizes are between the runs same much lower with the booster?

 

If yes, that could explain why booster shows higher Geekbench Scores, they may use not much memory blocks by their around 8 small bench tasks.

 


 

 

  • Like 1
Link to comment
Share on other sites

16 hours ago, mitch_de said:

So, if you also repeated the test are you sure also Bandwidths of the smaller transfersizes are between the runs same much lower with the booster?

 

If yes, that could explain why booster shows higher Geekbench Scores, they may use not much memory blocks by their around 8 small bench tasks.

yes results were comparable between tests.

Link to comment
Share on other sites

Don't use VGTab. It overrides AMD10000Controller's info.plist which is not what you want to happen. Also it doesn't offer an option for MinimumPWMLimit which is crucial for proper fan control.

Link to comment
Share on other sites

10 hours ago, Moviemakergr said:

On same build with Nitro+ Rx 5700XT works  on Gbench score on Cinebench score nothing change..

 

On cinebench cannot give any improvement it's a OpenGL Bench so in latest version 20 they removed it. Gbench test OpenCL and Metal most of the MacOs apps should use them not OpenGL.

Edited by Xanny
typos
Link to comment
Share on other sites

On 4/29/2020 at 1:04 PM, CMMChris said:

Don't use VGTab. It overrides AMD10000Controller's info.plist which is not what you want to happen. Also it doesn't offer an option for MinimumPWMLimit which is crucial for proper fan control.

 

Oh okay. I guess Vega 56 is out of luck

Link to comment
Share on other sites

On 4/19/2020 at 2:15 PM, CMMChris said:

New version 1.2 added in first post. Vega10 support has been removed (Vega 56, Vega 64, Vega Frontier).

 

Reason:
People have observed fan control being broken when using this kext like it was the case prior to the release of macOS 10.14.5 Mojave. You remember? 10.14.5 was the macOS release that fixed the fan issues with Vega 56, 64 and Frontier cards. But you might also remember that exactly with this release the Geekbench scores have become lower as well.

 

Thanks to the developers of Whatevergreen we now know that Apple disabled the loading of the SMU Firmware to PC graphics cards with this release of macOS. macOS does now check for the device property "ATY,EFIVersion" before uploading the SMU firmware to the graphics card. Adding it back causes the SMU firmware to be loaded again and is the main reason for the increase in Geekbench scores. Anyhow, it also causes the Vega10 fan bug to come back.

 

It really looks like Apple never fixed the Vega10 fan control issues which apparently are caused by a bug inside the SMU firmware. So they simply did a dirty hack, disabling the firmware for all PC cards and thus essentially crippling them in some way.

 

Since I cannot fix this issue and a silent card is more important than Geekbench scores, I decided to drop Vega10 support for RadeonBoost.

If you are pissed about the way Apple does handle bug fixing these days, open bug reports about this issue.

glad I found this thread and thanks for the explanation.  was driving me nuts that v1.0 RadeonBoost was working but my Vega 64 fans were not.  looking forward to RDNA 2.

Link to comment
Share on other sites

7 hours ago, CMMChris said:

 

With new bugs that will either never be fixed or with a delay of 1+ year(s)...

Hey man you gotta look forward to things :afro:

Link to comment
Share on other sites

On 4/21/2020 at 12:21 PM, Paksman said:

Hi, sorry if this is OT, i found a similar way from: https://github.com/acidanthera/WhateverGreen/blob/master/Manual/FAQ.Radeon.en.md. If you are using opencore with whatevergreen you can also inject gpu properties into config.plist, no need for kexts. It is working great and geekbench metal score skyrocketed from 39k to 66k+ on my rx 5700.

Screenshot 2020-04-21 at 13.14.32.png

Screenshot 2020-04-21 at 12.46.59.png

 

In OC, you do not add the PP, in the properties but just the property.

In your screenshot you show PP,PP_WorkLoadPolicyMask but it should simply be entered as PP_WorkLoadPolicyMask

Otherwise you will end up with TWO entries in the registry (one with PP, and another without) which what I found.

 

The PP, (and CFG,) prefix is used when doing the patches via an SSDT.

 

  • Like 1
Link to comment
Share on other sites

On 4/27/2020 at 11:19 PM, CMMChris said:

Yeah doesn't work for some cards. Apple's SMU firmware for Navi appears to be incompatible with some of the Navi cards for some reason. Nothing you or I can do about that unfortunately. Apple has to fix up their {censored} and make my kext obsolete :)

It won't panic if you use named framebuffer B). Do not copy ATY,Rom# unless you need it.

image.thumb.png.8c24e4fe6b02de7f6b09b0700c8dfebb.png

Edited by BRONNER
  • Like 1
  • Thanks 1
Link to comment
Share on other sites

2 hours ago, BRONNER said:

It won't panic if you use named framebuffer B).

Nice find my man. Works like a charm. Scores were ~48k and are now ~67k :thumbsup_anim:

 

Here's the complete entry for the lazy:

<key>PciRoot(0x0)/Pci(0x1,0x0)/Pci(0x0,0x0)/Pci(0x0,0x0)/Pci(0x0,0x0)</key>
			<dict>
				<key>@0,name</key>
				<string>ATY,Adder</string>
				<key>@1,name</key>
				<string>ATY,Adder</string>
				<key>@2,name</key>
				<string>ATY,Adder</string>
				<key>@3,name</key>
				<string>ATY,Adder</string>
				<key>ATY,Rom#</key>
				<string>113-D1820501-101</string>
				<key>Force_Load_FalconSMUFW</key>
				<data>
				AQ==
				</data>
				<key>PP_WorkLoadPolicyMask</key>
				<data>
				IA==
				</data>
				<key>device_type</key>
				<string>ATY,AdderParent</string>
				<key>name</key>
				<string>ATY,AdderParent</string>
			</dict>

ROM version will likely differ between all the cards (I have the reference XT) so make sure you add your correct one.

  • Like 2
Link to comment
Share on other sites

1 hour ago, Tiem said:

Nice find my man. Works like a charm. Scores were ~48k and are now ~67k :thumbsup_anim:

 

Here's the complete entry for the lazy:

ROM version will likely differ between all the cards (I have the reference XT) so make sure you add your correct one.

The "ATY,Rom#" is not required at all and you have to remove it unless your GPU's ROM provides correct value. It's an MSI-specific solution to avoid installer crash issue.

Edited by BRONNER
Link to comment
Share on other sites

 Share

×
×
  • Create New...