AMDGPU-PRO 16.60 Vulkan vs. Mesa 17.1-dev RADV Performance

[marek]

Michael marek is there any room for improvements RadeonSI OpenGL perf on Dota or is it at it's hardware/driver limits?

I'm asking this because the OpenGL performance is the same as AMDGPU-PRO's vulkan performance

We've not spent any time profiling and optimizing for DOTA, so I believe there's a lot of room for improvements. Threaded GL would be the first thing to try after verifying that it's CPU bound.

[bridgman]

Cool. Sounds like you would be well on your way to turning AMD into a billion dollar company (down from 5-6 billion).

After everything but the CPU business shuts down or shrinks into meaninglessness (since we would not be able to sell GPUs or APUs into any OEM without copy protection) what would you do for an encore ? Or is the idea that you would make enough money from CPU and GPU compute sales to subsidize GPU R&D while selling only into the channel to users who don't care about playing/viewing protected content ?

Anyways, sounds interesting. I look forward to seeing your business plan.

BTW the boss of my boss of my bosses is a her, not a him.

[bridgman]

Ahh, you're going to Make America Great Again. Excellent.

With respect, I'm not sure that walking away from the successful parts of American industry is the best way to accomplish that. Or is your idea that Intel/NVidia would make all the money and taxes from them (plus import taxes on avocados) would subsidize AMD R&D ?

I believe you would do better to start by adding products, so you can leverage existing designs while still keeping money flowing in rather than by walking away from existing products and markets as a first step.

Otherwise you end up burning through all your cash and more before your new products start paying the bills.

Of course if the plan is based on government subsidies and deficit funding then feel free to ignore everything I said. Any plan can work if you have enough of a sweetheart deal from a sufficiently large government.

[bridgman]

*bridgman scratches head, tries to reconcile "America first" with driving Intel & NVidia out of the USA.

I imagine you are talking about NVidia being dependent on fabs outside the USA, but Intel has R&D and fabs in the US today. What would stop them from selling ? Also what would stop NVidia from shifting to a US-based fab as well ?

I imagine everyone in the semiconductor business has contingency plans for this in place already.

[bridgman]

Ahh, OK. You need to be specific about that up front, since it has nothing to do with protectionism.

So you're talking about plans for the entire US semi industry, not just AMD. Presumably there will be discussions with Microsoft and the Hollywood content protection advocates as well ?

Without that, there will be a big black market in imported PCs that can play & view protected content, like prohibition all over again.

I'm not necessarily opposed to this of course... US prohibition was very good to Canada

[indepe]

There is a possibility that in a short time I will work for the BOSS of your BOSS of your BOSSES...

Interesting how many think they will be the voice in the ear of the boss of bosses...

[andrei_me]

Why should we negotiate with Microsoft and the Hollywood content protection advocates ?
People all around the world failed in doing this in the past. why should it be successful this time?
So no we will not negotiate with these people anymore.

"there will be a big black market in imported PCs that can play & view protected content, like prohibition all over again."

you have a simple Globalist-Bias thinking error... in short time there will not be any country left on planet earth not doing this. so from what country exactly do you want to smuggle into the US?

it is GAME OVER for these people.

Lost track if it is sarcasm, real, trolling or whatelse...

[Amarildo]

Why ís the Linux team so slow for releasing drivers? One new driver every 1.5 to 2 months, and not even all cards are listed. They could at least bring "non WQHL" drivers like there are for Windows.

[bridgman]

The release frequency is the same as Windows, other than point releases targeting specific new games, and for Linux the game-specific improvements tend to go into the open source stack more than the -PRO stack.

Releasing "non-WHQL" drivers would only save time if we were running the Linux drivers through Windows Hardware Quality Lab certification in the first place... which we are not, since none of the tests run on Linux. We could release "non-QA'ed" drivers but you have already said you don't want to take a chance on those.

As a general rule, the more frequently we release the less progress we make, since every release has overhead of its own (working with QA on test plans, dealing with false positive problems, updating packaging tools & testing/fixing etc..). It also means we are spending more time re-testing already-supported boards again and have less time for testing "new to the stack" HW which is what you want.

Remember that these releases are aimed at the slowly-moving enterprise-distro workstation market - they just happen to be attractive to consumer users in the short term as a way of getting OpenCL on newer distros (and Vulkan once more games appear). IIRC you are using a consumer card and faster-moving consumer distros, but you need better OpenCL than the open stack provides today ?

Two months seemed like a pretty good compromise, particularly since we are pushing most of the fixes & features into public repos frequently via the amd-staging-X.Y branches. More frequent releases would definitely interfere with adding new features & hw support.

[Amarildo]

Thank you for the detailed explanation.

I would take a chance on non-QA'ed drivers because all I need from the closed stack is OpenCL, all the rest suits me nice on the OSS stack.

I guess the problem is not only the pace at which the drivers are released, but the general (dis)organization of AMD. You had an OSS stack and a closed stack (Catalyst), the last one worked fine in all aspects for me. Then you decided to stop development of Catalyst in favor of AMDGPU-PRO (which is nothing more than the Catalyst stack with an opensource Kernel driver), but this left GCN 1.0 users stuck with ancient Catalyst for more than a year, because for some reason GCN 1.0 cards were not supported until now (lack of developers to support us, I guess).
OpenCL from Catalyst worked "OK" and I wouldn't mind using the entire stack to this day. The problem is that it doesn't work anymore, Blender simply will not render using Catalyst's OpenCL. OpenCL from 16.50 is buggy too, and I'm yet to test 16.60 today on a spare HD I just found.

I'll install Ubuntu tonight and see what happens.