-
Notifications
You must be signed in to change notification settings - Fork 350
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
pp_table issue: sclk dependant on voltage on latest kernels #433
Comments
and i am assuming we cant change the power limit as well? |
In pp_table, you can set power limit, wattage limit, or even target temp if you want to reduce power. Refer to the spreadsheet for location of those items. |
@rhlug were you able to control the power limit/voltgage for your vegas? Any instructions/steps you can direct me to? I want to reduce the power usage on my vega 56. I am running ubuntu 16.04, rocm 1.8.1. |
Oh yea. I have a couple small test rigs running Cryptonight-heavy. Vega56 runs ~1225 h/s @ ~100w (not TDP) and Vega64 runs 1325h/s @ ~105w (not TDP). I dont have any Cryptonightv7 rates to share.
I run Ubuntu 18.04 with 4.17.0-rc2-180424-fkxamd kernel built from fxkamd/drm-next-wip ROCT-Thunk. I apply pp_tables that set them at 1365mhz core (900mv) and 925mhz memory. My last card is XFX reference, and doesnt like 925mhz.
I have 64 bios on these ref 56's also, and I can push them all at 1095mhz memory no problem, so not sure why it doesnt like 900+mhz on 56 timings. I can push the first 3 at 945mhz no issue. I find there is no speed benefit to running 64 bios on my vega56 in terms of cryptonight*. If I were mining eth, I'd be flipping to 64 bios, and running 1095mhz for sure. They can actually go higher if I up the SoC. But I dont like heat in the summer, so not mining ethash. |
@rhlug do you mind giving me some pointers on how to build a 4.17.0-rc2-180424-fkxamd kernel with ROCT-Thunk |
something like this.. (editing, sorry but i realized i gave wrong url for kernel)
Thats the kernel part of it. The Thunk
|
@rhulg thanks that helps a lot. I read a post where you mention you have amdgpu-pro 18.20-579836 without DKMS, how did you add this? |
@TheKnightCoder When you look at how stack is build you have Set of Language Runtimes, Usermode Driver and Kernel mode driver. For AMD all driver start on linux with AMDGPU base kernel mode driver which is what is upstreamed in the linux kernel. The each driver can have different UMD layer. AMDGPUpro use PAL + LLVM to HSAIl -> HSAIL Finalizer -> Shader Compiler aka SC. ROCm user mode is ROCr ( HSA System Runtime based ) with THUNK layer OpenCL as core language library is the same for ROCm and AMDGPUpro what different is VDI layer mappings. AMDGPUpro goes to PAL and ROCm goes to ROCr. Second change is they run on two different Compilers But use a common Frontend. AMDGPUpro still used proprietary Shader Compiler which is also used for OpenGL. ROCm uses native LLVM code generator. When you see performance delta it can be optimization flag you need to set like shutting off DENORMS, or they could be need to address optimization maturity issue native LLVM compiler. It newer. But you also have all the source to it so you can see what going on. KCL in DKMS is issue right now we discussing with Linux team way to fix it, also make sure they have compatibily of what they upstream with the ROCm thunk. This is the current issue. |
@gstoner thanks for the explanation it's making a lot of sense now. But if ROCm and AMDGPUpro are two different usermode drivers then why does some AMDGPUpro mention they include ROCm 1.6? Specifically the page says: |
Which version of AMDGPUpro.. 18.xx series is when the switch happened, Older AMDGPUpro used ROCm aka ROCr OpenCL runtime in 17.50 with the LLVM to HSAIL compiler.
On Jul 7, 2018, at 1:50 PM, TheKnightCoder <notifications@github.com<mailto:notifications@github.com>> wrote:
@gstoner<https://github.com/gstoner> thanks for the explanation it's making a lot of sense now.
But if ROCm and AMDGPUpro are two different usermode drivers then why does some AMDGPUpro mention they include ROCm 1.6?
Specifically the page says:
Package Contents
AMDGPU-Pro Driver
ROCm Platform 1.6 in supported distributions
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub<#433 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AD8DuSzcXmLPBaEvzTf23NBb8BBqrIIGks5uEQMIgaJpZM4Um0ET>.
|
With the fkxamd kernel, the amdgpu-dkms package will install, it will just fail to install the dkms module. But thats okay, because you dont need it. All the kfd is in the kernel.
Right now, just amdgpu-pro 18.20-606296. I'll probably be switching to rocm 1.9 when its out. |
So a question . If you are starting from base 18.04 with 4.17rc2 say. With no rocm or amdgpupro installed. Are you saying you can just build and install the 2 drivers and it will work with opencl etc. Or do you have to add anything before or after. Really really appreciate a full list of things on base system just installed and kernel updated. With exactly what you have to build and install to get opencl rocm interface working |
|
This is only for GFX8. and older GPU's, it is not used for Vega10 |
i am so confused. I just want to install a kernel and be able to run tdxminer which is dependent upon rocm and also undervolt my vega64's Is there any way i can do that? |
If you are using 18.04, but you want to use a kernel newer than the 4.15 that comes with current versions of Ubuntu 18.04, then you should look into using ROCm with our upstream Linux driver support. To do this, you can follow these directions:
Note that if you do not want to follow these directions manually, you should also be able to run the Ubuntu 18.10 installation scripts in the Experimental ROC project to automatically perform these steps. The Ubuntu 18.10 installation directions use the upstream kernel driver, and basically every other step is the same as installing Ubuntu 18.04. So they should work if you're using a post-4.15 kernel on Ubuntu 18.04. Note that if you are using a Vega 10 GPU, you will need a post-4.17 kernel to have the required software support. If you want to try overclocking, undervolting, etc. on your GPU, you can look into using new versions of You can also use these tools to potentially set a higher maximum power cap on your GPU, as requested here by @heavyarms2112 and @Sayyiditow. This can be done with @TheKnightCoder if you are interested in building custom versions of any of the ROCm software stack, please check out the Experimental ROC project, where we now have build scripts for various Linux distributions for all of the software projects in ROCm. |
There may be a new issue around pp_table on latest kernels. I wanted to record it here in case it flows through to amdgpu-pro or rocm via dkms being loaded.
I was testing powerplay in kernel drm-next-4.19-wip.
I setup a pp_table on a couple vega64 cards for 1390mhz/1090mhz/900mv
Both cards say they are running those clocks (via /sys/class/drm/*)
However, getting data from /sys/kernel/debug/dri/*/amdgpu_pm_info, its different.
I dont have any idea where gpu #0 gets 937mv. Not from pp_table, thats for sure.
I cant get to 1390mhz core unless I up voltage over 975mv. Here is 925mv, 950mv, 975mv
Reverting back to 4.17.0-rc2-180424-fkxamd, what I set in pp_table is really happening
The text was updated successfully, but these errors were encountered: