The AMDGPU Info/News Thread

Does anyone have idea how to “tune” current scheduler in amdgpu, so it goes agressive, as for now to get good performance, i have to put “high” in '/sys/class/drm/card0/device/power_dpm_force_performance_level’
any workarounds maybe? (like with ondeman trasholds or something).

The card should be able to tune itself automatically, you shouldn’t have to do that work around. That’s been a feature for a long time now.

That particular work around was needed back in earlier kernel days when there was a low-power bug affecting Hawaii cards; so if they ever went into a low-power state they would halt the system. It was fixed in kernel 4.10 IIRC.

Just make sure you’re running that or higher.

Yeah install the newest kernel. The driver does that for you. Its nice not needing to OC.

I have been waiting for the overclocking to work for a polaris card. Maybe 4.15 with allow you to do it. On 4.13 I could change some settings but the performance was slower. At the moment I cant change it it with permission denied.

https://linuxconfig.org/overclock-your-radeon-gpu-with-amdgpu

I spy with my little eye raven ridge firmware in disguise

update-initramfs: Generating /boot/initrd.img-4.13.0-32-generic
W: Possible missing firmware /lib/firmware/amdgpu/raven_gpu_info.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_rlc.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_mec2.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_mec.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_me.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_pfp.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_ce.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_sdma.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_vcn.bin for module amdgpu
update-initramfs: Generating /boot/initrd.img-4.13.0-31-generic
W: Possible missing firmware /lib/firmware/amdgpu/raven_gpu_info.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_rlc.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_mec2.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_mec.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_me.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_pfp.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_ce.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_sdma.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_vcn.bin for module amdgpu

update-initramfs: Generating /boot/initrd.img-4.13.0-25-generic
W: Possible missing firmware /lib/firmware/amdgpu/raven_gpu_info.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_rlc.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_mec2.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_mec.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_me.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_pfp.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_ce.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_sdma.bin for module amdgpu
W: Possible missing firmware /lib/firmware/amdgpu/raven_vcn.bin for module amdgpu
2 Likes

I noticed this too the other day

Moved to correct Thread

https://wccftech.com/exclusive-amd-navi-gpu-roadmap-cost-zen/

Remember when Lisa said that ‘7nm GPUs will be arriving for gamers’? Well, notice she never said 7nm Vega will be arriving for gamers – the GPU in question was Navi. Navi 10 to be more specific, and from what I know, this will not be a high-end part. It looks like NVIDIA won’t be facing any competition on the gaming side of things from Radeon even in 2019.

Well… I mean… time enoguh to save up for some sweet NextGen ™ GPUs…

1 Like

I don’t know if there is any truth to it, but some software engineer types at computex had some very interesting stories to tell. They said that there is hardware that is not activated on vega, and that there were a lot of things in the silicon that are supposed to do multiple steps at once but they didn’t figure out the corresponding software side of the implementation.

1 Like

I wonder if thats something we could hack out of the drivers to see what it does :stuck_out_tongue:

I figured as much, but did they perhaps indicate “not activated” as in fused off in hardware, or driver/firmware development dependent?

Not fused off, literally just how to translate what the game wants to do into what stuff can be done in one cycle on silicon.

2 Likes

just to get it right: there is tech in vega that is not active that is supposed to figure out how to do games in cycles and then parallels those?

I was told that there are instructions or operations available in the silicon that is not really being used in the drivers, that they may never get to use, that was expected to speed up performance a lot over where they are now. But there is some issue, or something. This wasn’t from anyone inside amd, but someone I met that has the software engineering expertise to make the call.

Is it a question of the feature was mis-implemented in silicon (oh, we implemented add-multiply-move when we should have implemented decrement-multiply-move, whopise!))? or a software engineering problem? Or perhaps the approach they wanted to use was encumbered by some type of patent? No idea.

How insane would it be if it were some patented method of doing something that exists in silicon, but they aren’t using for fear of a lawsuit. That’d be extra crazy, for sure. But we didn’t really talk about it from that aspect.

5 Likes

is there any chance that one can figure out what kind of instructions and try to access those or is it not possible to activate those instructions?

Hidden instructions in a GPU is likely very difficult to use because you are usually talking with the driver, and the driver isn’t using the cool instructions.

It is not impossible if they actually are properly implemented in the silicon, but it would be very difficult to find and access.

Outside proprietary secrets. It would be cool if they wrote a white paper on hardware instructions / features not implemented and invited open source / creative ways to use them.

I guess that is way too much of a mine field to actually expect to happen. The first thing done would be mining or cracking something.

Its HSA tech I guarantee it.

excellent. so that means potential driver or software improvement. which is what i’ve been kinda betting on. not relying on it, but guessing things will improve with time. :blush:

i suspect / hope it may just be software. vega 7nm sees suggested 35 percent gain from amd slides.

smaller process. if it was a new hardware revision to go to 7nm presumably they would have fixed any obvious hardware issue…?