The RX Vega 56 Owners Thread

erazor51 · 15 Sep 2017 at 11:58

SpudMaster said:
Just ran Superposition 1080p extreme on my 56 with 64 bios.
I got 4022 with card set at 1700mhz and 1100mhz HBM. Seemed to maintain a boost of between 1650 and 1680mhz.
4022 seems a little low. I get over 7600 gpu score in Timespy

4022 is low for those settings. Mine set to default volts, PL +50, HBM 920 and I got around 4250~ with frequency peaking at 1480mhz.

With undervolt my card only peaks to around 1300mhz and i get a score of 3649, which I hope I can fix in the future.

old gamer · 15 Sep 2017 at 11:59

SpudMaster said:
Just ran Superposition 1080p extreme on my 56 with 64 bios.
I got 4022 with card set at 1700mhz and 1100mhz HBM. Seemed to maintain a boost of between 1650 and 1680mhz.
4022 seems a little low. I get over 7600 gpu score in Timespy

yep seems a bit low, I'm getting some odd results myself, I get a higher score with my 1700 running stock than I do with it overclocked, spent all morning adjusting settings and getting nowhere

lowdog · 15 Sep 2017 at 13:27

Results are weird for me, 56 flashed to Sapphire 64 AIO Liquid bios;

Ryzen 1800X @ 3.9GHz with 1.375V and 32Gb Ram @ 3333MHz 14/14/14/34

+50 PL / S6 core 1637 and 1200mv - S7 core 1722 and 1250mv - mem 1050 and 950mv = constant/stable core speed of 1720MHz (max for my chip) with a roughly 340W power draw on the card and 580W total system at the wall.

+50 PL / S6 core 1627 and 950mv - S7 core 1712 and 950mv - mem 1050 and 950mv = constant/stable core speed of 1595/1600MHz with 200/210W power draw on the card and roughly 400W total at the wall.

TheRealDeal · 15 Sep 2017 at 13:46

I am struggling to see the weird part. Run 1 is using higher voltage with higher clocks. Run 1 is faster and uses more power.

Run two is uses much lower voltage and lower clocks. Run two scores lower and uses less voltage at the wall.

To me this looks normal or is there something i am missing.

T.Kay · 15 Sep 2017 at 13:50

SpudMaster said:
Just ran Superposition 1080p extreme on my 56 with 64 bios.
I got 4022 with card set at 1700mhz and 1100mhz HBM. Seemed to maintain a boost of between 1650 and 1680mhz.
4022 seems a little low. I get over 7600 gpu score in Timespy

Could you post ur watt man settings, try trying on hbcc for the superposition benchmark really helps the score.

FelixB · 15 Sep 2017 at 15:09

Superposition benchmarks are very sensitive to the stability of the overclock and drop quickly if the GPU shows any sign of throttling. If you are sticking to the default Vega 56 power limits then a 1020mV undervolt and a 1% overclock may give the best result. Flashing to Vega 64 would relax the power limit, but it might be better to take a direct route and change the power play table in the registry. Either way, you want good cooling to get good results when raising the power limit.

calken · 15 Sep 2017 at 15:10

SpudMaster said:
Just watched this. Should make 56 owners happy

I watched this yesterday and it does suggest that something is clearly un-optimised.

In regard to supply, I recon they can get about 110 dies per 300mm wafer and if dedicating 10% 14nm capacity to Vega, that's around over 500,000 dies a month with a 75% yield. Of course, there are some real finger in the air numbers here!

Is it comfortable to suggest over 1 million Vega's have been sold? Who knows.

Refs:
https://en.wikipedia.org/wiki/GlobalFoundries
http://www.silicon-edge.co.uk/j/index.php/resources/die-per-wafer
https://www.reddit.com/r/Amd/comments/6krrjj/raja_spills_the_beans_on_vega_10_die_size_484mm²/

FelixB · 15 Sep 2017 at 15:31

Using Vega on Linux
===================

This is for the serious enthusiast. Proceed at your own risk and do not expect help.

Vega is VERY particular in its requirements under Linux, so you will need to start from scratch and install a new OS on the bare metal. This is inherently risky, so it is best to reboot the computer with a single drive in it that you are happy to have reformatted. Before you begin, you need to work out how to select different UEFI boot table entries from your mother board BIOS, so that you can recover from the changes that installing the new OS will make to the boot table.

You will need at least 25G of unallocated disk space available for the installation.

Download the Ubuntu 16.04.3 installation iso file from:

http://releases.ubuntu.com/16.04/ubuntu-16.04.3-desktop-amd64.iso

Use Ruby to create a UEFI bootable usb stick containing this iso. Boot and install.

The newly installed system will use unaccelerated drivers. You will need to install accelerated drivers before doing anything useful.

You can EITHER install the proprietory (amdgpu-pro) accelerated drivers OR the open source (amdgpu) accelerated drivers. If you wish to switch between them then you must remove the installed drivers completely and reboot before installing the other drivers.

Using the AMD proprietory Vega drivers (amdgpu-pro)
---------------------------------------------------

The official AMD documentation is available at:

http://support.amd.com/en-us/kb-articles/Pages/AMDGPU-PRO-Install.aspx

Carefully follow their instructions.

Using the open source Vega drivers (amdgpu)
-------------------------------------------

Download an amdgpu compatible ubuntu linux kernel from:

https://github.com/M-Bab/linux-kernel-amdgpu-binaries

(I used a kernel version from his site (4.13.1) that he has since removed on grounds of instability. I have yet to try his latest version, but it should work.)

Download the Vega firmware:

https://github.com/M-Bab/linux-kernel-amdgpu-binaries/blob/master/firmware-radeon-ucode_1.40_all.deb

Install the kernel and firmware by executing this at the command prompt ("terminal" in linux speak):

Code:

sudo dpkg -i linux-image-4.12.12+_17.09.12.AMDGPU.UBUNTU_amd64.deb
sudo dpkg -i firmware-radeon-ucode_1.40_all.deb

Reboot the system.

Install the amdgpu drivers from a stable repository by executing:

Code:

sudo add-apt-repository ppa:paulo-miguel-dias/pkppa
sudo apt-get update
sudo apt-get upgrade
sudo apt-get install libegl1-mesa libgbm1 libgl1-mesa-dri libgl1-mesa-glx libglapi-mesa libwayland-egl1-mesa libxatracker2 libdrm-amdgpu1 libdrm-intel1 libdrm-nouveau2 libdrm-radeon1 libdrm2

Reboot.

You will get error dialogs ("System Program Problem Detected") flashed up on the screen while using the open source drivers. These do not mean anything serious, but you will have to press cancel on them repeatedly before you can do anything else.

You can remove the amdgpu drivers and compatible kernel by executing:

Code:

sudo apt-get install ppa-purge
sudo ppa-purge ppa:paulo-miguel-dias/pkppa
sudo apt-get update
sudo apt-get upgrade
sudo apt autoremove

Reboot using "Advanced Options for Ubuntu" and then "Ubuntu, with Linux 4.10.0-33-generic" and then execute:

Code:

sudo apt-get remove firmware-radeon-ucode_1.40_all
sudo apt-get remove linux-image-4.12.12+_17.09.12.AMDGPU.UBUNTU_amd64

Any mistake in installing or removing the drivers might make the system unstable. Reboot and select "Advanced Options for Ubuntu" and then "Linux 4.10.0-33-generic (recovery mode)" and then "Repair packages". This usually gets things back into a stable state.

Overclocking, etc.
------------------

I have yet to explore overclocking Vega on linux. Linux does not have Wattman. Some equivalent functionality (like percentage overclocking) is available by writing values to the powerplay entries under the /sys/ directory. Other functionality (like undervolting or relaxing power limits) may require direct patches to the /lib/firmware/amdgpu/vega10_* powerplay files and/or patches to the kernel sources.

In the meantime, here are some Unigine benchmark scores (comparing raw Vega 56 performance on linux against overclocked performance on Windows):

Code:

Score        : amdgpu : amdpro : win GL :  DX11
------------- -------- -------- -------- --------
Heaven 1440p :  1384  :  1229  :  1561  :  1933
Valley 1440p :  2101  :  2229  :  2547  :  2568
Superpos  4K :  4359  :  2859  :  4247  :  5842
Superpos  8K :  1804  :  1401  :  1871  :  2540

Average FPS  : amdgpu : amdpro : win GL :  DX11
------------- -------- -------- -------- --------
Heaven 1440p :  54.9  :  48.8  :  62.0  :  76.7
Valley 1440p :  50.2  :  53.3  :  60.9  :  63.5
Superpos  4K :  32.6  :  21.4  :  31.8  :  43.7
Superpos  8K :  13.5  :  10.5  :  14.0  :  19.0

Notably, the open source linux drivers are usually better than the closed source linux drivers and sometimes better than the overclocked OpenGL drivers on Windows. However, the overclocked Windows DX11 drivers always come out best.

These comparisons show how important the drivers are in determining performance and how rapidly the open source drivers are catching up. The testing at Phoronix tells a similar story, but adds some icing on the cake: following optimisation on the open source amdgpu drivers, the stock Vega 56 performs as well as a GTX 1080 Ti on Bioshock Infinite at 4k.

T.Kay · 15 Sep 2017 at 16:27

FelixB said:
Superposition benchmarks are very sensitive to the stability of the overclock and drop quickly if the GPU shows any sign of throttling. If you are sticking to the default Vega 56 power limits then a 1020mV undervolt and a 1% overclock may give the best result. Flashing to Vega 64 would relax the power limit, but it might be better to take a direct route and change the power play table in the registry. Either way, you want good cooling to get good results when raising the power limit.

Oh okay so is this why I've been getting lower scores with a higher memory and core overclock?

FelixB · 15 Sep 2017 at 17:09

T.Kay said:
Oh okay so is this why I've been getting lower scores with a higher memory and core overclock?

Quite possibly. As Steve says in the video above, Vega will automatically throttle if the power, thermals, etc. start hitting their prescribed limits. The throttled rate may be 10-20% lower than the peak and it will take time to return to stability. This can be a major bite out of your performance. You can undervolt to keep below the power and thermal limits, but you can not escape them. I see better Superposition performance at 1% overclock than stock, but anything greater and the scores fall.

I suggest that you start at stock and work up slowly, keeping a close eye on fluctuations in the clock speed.

SpudMaster · 15 Sep 2017 at 17:20

My card is water-cooled and didn't go over 39℃ during my run.

Will post my wattman settings as requested later tonight when home from work, wife/kids sorted out etc

erazor51 · 15 Sep 2017 at 17:34

FelixB said:
Quite possibly. As Steve says in the video above, Vega will automatically throttle if the power, thermals, etc. start hitting their prescribed limits. The throttled rate may be 10-20% lower than the peak and it will take time to return to stability. This can be a major bite out of your performance. You can undervolt to keep below the power and thermal limits, but you can not escape them. I see better Superposition performance at 1% overclock than stock, but anything greater and the scores fall.

I suggest that you start at stock and work up slowly, keeping a close eye on fluctuations in the clock speed.

Can you share your wattman settings with us please? thanks

lowdog · 15 Sep 2017 at 20:57

TheRealDeal said:
I am struggling to see the weird part. Run 1 is using higher voltage with higher clocks. Run 1 is faster and uses more power.

Run two is uses much lower voltage and lower clocks. Run two scores lower and uses less voltage at the wall.

To me this looks normal or is there something i am missing.

Weird may have been the wrong word.....I suppose I'm just surprised that on the second run this chip can sustain 1600MHz core with the 950mv settings but for the extra 120Mhz for 1720MHz core I need to go to 1250mv.....quite a jump in mv for the extra 100+ MHz

SpudMaster · 15 Sep 2017 at 23:41

T.Kay said:
Could you post ur watt man settings, try trying on hbcc for the superposition benchmark really helps the score.

Just ran the Superposition 1080p Extreme benchmark again but with HBCC enable and got a score of 4820 which looks much better

Are you guys using HBCC all the time? I didn't think it would make any difference unless you used more than the 8gb HBM

My WattMan setting are:
State 6: 1637mhz, 1150mV
State 7: 1702mhz, 1200mV
HBM 1100mhz, 1050mV
Power Limit: 50%
Running the 64 Air bios.

Haven't really played about with the voltages yet so they are just as the defaults of the 64 bios.
During the Superposition benchmark my card maintained a boost of 1660mhz and during games of tends to go between 1650mhz and 1680mhz

erazor51 · 16 Sep 2017 at 00:05

SpudMaster said:
Just ran the Superposition 1080p Extreme benchmark again but with HBCC enable and got a score of 4820 which looks much better
Are you guys using HBCC all the time? I didn't think it would make any difference unless you used more than the 8gb HBM

My WattMan setting are:
State 6: 1637mhz, 1150mV
State 7: 1702mhz, 1200mV
HBM 1100mhz, 1050mV
Power Limit: 50%
Running the 64 Air bios.

Haven't really played about with the voltages yet so they are just as the defaults of the 64 bios.
During the Superposition benchmark my card maintained a boost of 1660mhz and during games of tends to go between 1650mhz and 1680mhz

Nice scores! What is your fan and temp set to?

HBCC is always on as it doesn't give any negative impact.

SpudMaster · 16 Sep 2017 at 00:51

erazor51 said:
Nice scores! What is your fan and temp set to?

HBCC is always on as it doesn't give any negative impact.

Fan and temps are at defaults as the card has a full cover EK block on it

lowdog · 16 Sep 2017 at 00:56

SpudMaster said:
Just ran the Superposition 1080p Extreme benchmark again but with HBCC enable and got a score of 4820 which looks much better
Are you guys using HBCC all the time? I didn't think it would make any difference unless you used more than the 8gb HBM

My WattMan setting are:
State 6: 1637mhz, 1150mV
State 7: 1702mhz, 1200mV
HBM 1100mhz, 1050mV
Power Limit: 50%
Running the 64 Air bios.

Haven't really played about with the voltages yet so they are just as the defaults of the 64 bios.
During the Superposition benchmark my card maintained a boost of 1660mhz and during games of tends to go between 1650mhz and 1680mhz

You sure that was the 1080p Extreme preset and not the High preset???....score seems a little too high for the extreme preset considering your clocks.

T.Kay · 16 Sep 2017 at 01:00

Th

lowdog said:
You sure that was the 1080p Extreme preset and not the High preset???....score seems a little too high for the extreme preset considering your clocks.

That score is correct, just got 4551 at 1558mhz core and 1100 memory.

lowdog · 16 Sep 2017 at 01:08

T.Kay said:
Th

That score is correct, just got 4551 at 1558mhz core and 1100 memory.

WOW....best I can get with 1800X @ 3.9Ghz ram @ 333MHz and Vega with 1730MHz solid core speed and mem 1050Mhz is 4380 at 1080p Extreme preset (HBCC makes no difference) ......Vega ram speed between 950 - 1050 make negligible difference so I'm just wondering how you can get 4551 with a 1558 core speed :confused:

If you look here you will see no one is getting those scores with Extreme preset regardless of clocks. https://docs.google.com/spreadsheets/d/1qBRTNCluCB6wTBqCyzbWjIPLa2d2BKL-mPCc48-MWKg/edit#gid=0

T.Kay · 16 Sep 2017 at 01:11

lowdog said:
WOW....best I can get with 1800X @ 3.9Ghz ram @ 333MHz and Vega with 1730MHz solid core speed and mem 1050Mhz is 4380 at 1080p Extreme preset (HBCC makes no difference) ......Vega ram speed between 950 - 1050 make negligible difference so I'm just wondering how you can get 4551 with a 1558 core speed

Have you tried maxing out the hbcc memory size? If you guys check the superposition benchmark thread some vega 64 get near 1080ti scores!!