Post OS X Luxmark OpenCL Benchmarks

[FernandoCoverdale]

To prevent it from throttling, you need to improve the cooling of the video card.

at the software or hardware level?

 

[pastrychef]

at the software or hardware level?

Either or both. Cool it any way you can. You can ramp up the fans in VGTab and/or change the cooler.

The Vegas have a lot headroom as long as you can keep it cool.

 

[jaymonkey]

The values of my Vega are below yours, I don't know how to solve it.

@FernandoCoverdale,

As @pastrychef commented, my Vega 64 card is the liquid cooled edition, these cards use binned GPU chips to ensure that they can reach the high default clock rate of 1.75 Ghz. This is really about the maximum anyone can expect to get a good Vega chip to run at without resorting to modifications of the boards power circuits which is not recommend and will result in shorting the life of the GPU.

Do you recommend any extra adjustments? thank you

I haven't touched the core/die clock rates on my card as I very much doubt that it would be stable beyond 1.75 Ghz, however you may be able to get a slight boost in performance by increasing the core/die clock rate on your card, but do it in small steps and test for stability and monitor temps by running a GPU stress test after each change.

I used a SoftPowerPlayTable to bump the HBM clock to 1100 Mhz on my card which resulted in a bit of a performance boost, most Vegas should allow a HBM clock rate of 1000 Mhz ... possibly even higher depending upon the silicon lottery, again make the changes in small steps (no more than +50 Mhz each time) and check for stability and temps.

You'll probably see more of a boost in performance with the HBM clock rather than pushing the core/die clock. With careful adjustment of the SoftPowerPlayTable most Air Cooled Vega 64's can reach a Luxmark score around 32,500.

WARNING: You should be aware that bumping the Die and/or the HBM clock rates on a Vega GPU will result in the card running hotter and most air cooled Vega's already run hot at stock clocks, so unless you can keep your card cool it is not recommended. As my Card is liquid cooled it never get above 60 Degrees under a full load stress test even with the HBM 1100 Mhz overclock in place.

If the card draws more power than the VBIOS allows then it will automatically default to the low base clock settings (usually 800 Mhz Core and 945 Mhz HBM) so you need to monitor those during testing after making changes.

You can use the VGTab Utility that @pastrychef posted a link to above to generate a kext, however it does cause conflicts with Lilu + WhatEverGreen so if you have them installed you can inject the custom SoftPowerPlayTable via Clover using the method detailed in this guide :-

[GUIDE] Injection of AMD Vega Power and Fan Control Properties

Injection of Device Properties for Power and Fan control Last updated 14th May 2019 Update: MacOS Mojave 10.14.5+ includes a major overhaul of the Vega GPU drivers which should resolve all the power and fan control issues with most Vega cards and thus should make the methods outlined in this...

www.tonymacx86.com

Cheers
Jay

 

[jaymonkey]

jaymonkey's Vega 64 is liquid cooled and runs at 1.71GHz.

@pastrychef,

Slight typo on my part, the stock Max Clock for LC Vega 64 is 1.75 Ghz not 1.71 Ghz
Not sure why i typed it wrong in that post, corrected now.

Cheers
Jay

 

[oddless]

With a Vega 56, I was only able to get to a 950MHz memory clock (stock is 800MHz). Pushing to 1000MHz wasn't possible even though temps were low (~70ºC). It did bump up my LuxMark score up a little, but not a whole lot (less than 400 pts). Odd thing is I get better Geekbench compute results without the overclock

 

[shasen1235]

With a Vega 56, I was only able to get to a 950MHz memory clock (stock is 800MHz). Pushing to 1000MHz wasn't possible even though temps were low (~70ºC).

Check your HBM2 vendor with GPU-Z and see whether it's from Hynix or Samsung.
If it's from Hynix then sorry you're most likely not gonna get much OC from it.
But if it's from Samsung, 1050MHz+ is mostly guaranteed.

And if you want to OC your card to it's full potential, flash a Vega 64 vBios.
This will allow your HBM2 voltage go up to 1.35v(compared to 56's 1.2v)and more OC headroom.

 

[oddless]

Check your HBM2 vendor with GPU-Z and see whether it's from Hynix or Samsung.
If it's from Hynix then sorry you're most likely not gonna get much OC from it.
But if it's from Samsung, 1050MHz+ is mostly guaranteed.

And if you want to OC your card to it's full potential, flash a Vega 64 vBios.
This will allow your HBM2 voltage go up to 1.35v(compared to 56's 1.2v)and more OC headroom.

I don't have a copy of Windows and not really interested in installing it. Pretty sure there's a high probability the memory is made by Hynix as that seems to be what most 56s shipped with. And the fact that I can't get to 1000. =] TBH, I'm happy with stock performance. It's a little better than the 5 year old GTX 6xx 1GB card I was using in the last build

 

[shasen1235]

I don't have a copy of Windows and not really interested in installing it. Pretty sure there's a high probability the memory is made by Hynix as that seems to be what most 56s shipped with. And the fact that I can't get to 1000. =] TBH, I'm happy with stock performance. It's a little better than the 5 year old GTX 6xx 1GB card I was using in the last build

Actually from what I've experienced yours is most likely from Samsung.

I suppose you haven't flashed the vBios and yet you've already pushed it to 950

Normally Hynix memories in 56 with 64 vBios's higher voltage is capped around 950

Unless you won the "Hynix lottery" or there's no way a Hynix memory can be clock that high with stock vBios.

So if you want to push it further, there's still some headroom in your card

 

[jaymonkey]

With a Vega 56, I was only able to get to a 950MHz memory clock (stock is 800MHz). Pushing to 1000MHz wasn't possible even though temps were low (~70ºC). It did bump up my LuxMark score up a little, but not a whole lot (less than 400 pts). Odd thing is I get better Geekbench compute results without the overclock

@oddless,

If you get a better Geekbench score without the OC then it would suggest that the GPU is automatically under clocking when under load ..

Suggest you monitor the die and HBM clocks when running a LuxMark test and see if they fluctuate (they should not).

Luxmark will always peg the GPU die and Memory clocks when running a test.
GeekBench compute test runs a series of chained tests many of which do not peg the GPU.

If you observe clock fluctuations when running a LuxMark test then you either need to lower the OC or give it a little more voltage, but if the card goes over a preset power limit (depends on VBios) then it will also automatically under clock.

Also make sure you have these Device Properties for the Vega as well as the PP_PhmSoftPowerPlayTable

                <key>PP_DisableClockStretcher</key>
                <integer>1</integer>
                <key>PP_DisablePowerContainment</key>
                <integer>1</integer>
                <key>PP_DisableULV</key>
                <integer>1</integer>

Cheers
Jay

 

[oddless]

Thanks @jaymonkey. I adjusted the memory voltage to 975MHz @ 1275mv and bumped the core clock to 1650MHz and undervolted to 1125. That seems to be stable and LuxMark just returned a 31,939. When I have more time, I'll see if I can get it to 1GHz under the same voltage settings.


UPDATE: I ran a second test after quitting some running processes that start at login and surpassed the 32k mark

But I'm still stumped as to why Geekbench 4 Compute benchmarks are so much lower than stock. I might try bumping up earlier core/memory clock rates to see if that helps. The GPU and memory clock rates are pretty much dormant during these tests.