AMD's Vulkan Driver Will Only Work With The AMDGPU Kernel Driver

[sarmad]

Seems a pretty bad news.

This is a clear message to game devs 'you need to support gl if you want Linux customers'... Thus slowing down Vulkan adoption significantly...

Not really. There is currently no Steam Machine with AMD chipset, so developers won't really care much about older AMD gpus. So, this is not bad news for Vulkan, it's bad news for people with older AMD gpus, myself included.

[agd5f]

You inflate it a lot. We all now that Vulcan needs only SM5a and classic Compute Shaders. The rest can be done on software or you made it up. Fake reasons to exclude HD5000-6000 series. But don't thing that will go your way, a lot of people can program (my self included).

You could do theoretically do vulkan on pre-GCN parts, but it you wouldn't gain much over existing APIs. Pre-GCN hardware does not have independent compute pipes, so compute and graphics could not be submitted asynchronously and executed in parallel. Shader resources are fixed so you can only have a limited number of resources in flight at any given time and you'd have to explicitly bind the resources at execution time. If someone wants to work on vulkan support for older parts, I certainly won't stop you, just don't expect much of a gain compared to older APIs.

[theriddick]

AMD seriously will need to support 200/300 series of cards at a BARE MINIMAL, if they just don't then AMD! LISTEN! Expect significant market share drop to continue throughout 2016. I don't think AMD can afford to lose what little Linux/Vulkan market share they have atm.

So it would be a VERY BAD move to only support GCN 1.2 cards in the new drivers which are DESPERATELY needed for Linux!

[M@yeulC]

A bit sad, but I'm OK with that. My upgrade path has been the following :

HD 4850 : very impressive card. Held quite some time against the newest games.
Then... Was not enough for some demanding games on Linux. Moreover, I've had bad luck with catalyst, thus witching to the OSS drivers, which was a really pleasant experience for me.
Enter the 6870... Best of its competitors from Michael's Open source benchmark at the time. I got one for € 50 online.
As of today, it's a bit difficult to play every game maxed out on Linux, in part due to aging hardware (I have been waiting for zen to see how they compare to intel). But I've been closely following AMD's unified driver strategy, and Vulkan driver developments. Mantle required GCN 1.1 IIRC, and AMD told us the Vulkan spec might allow it to be compatible with GCN 1.0, but we didn't get any confirmation (at least on the top of my head).
So... I'm just waiting for this strategy to come into fruition before buying any new graphics hardware (including upgrading my 1650x1080 screen, at full HD there are 309 600 more pixels to compute, roughly ~18% more compute power required).
So... It's a little bit like chess: with a bit of patience, carefully planned decisions and such, I don't see why everyone couldn't enjoy their Vulkan-ready GPU in some time (heck, the spec isn't even out).

That said... I suck at chess. So there are probably people out there with better strategies. Oh, and I am a student, which means (very) low income, so not much to spend on shiny new hardware.

You could do theoretically do vulkan on pre-GCN parts, but it you wouldn't gain much over existing APIs. Pre-GCN hardware does not have independent compute pipes, so compute and graphics could not be submitted asynchronously and executed in parallel. Shader resources are fixed so you can only have a limited number of resources in flight at any given time and you'd have to explicitly bind the resources at execution time. If someone wants to work on vulkan support for older parts, I certainly won't stop you, just don't expect much of a gain compared to older APIs.

I think the question might be one of marketshare too: as a developer, you need to have the biggest reach on your market when choosing which API to use (the same could be said about OSes, of course, but that's one less problem if every card supports the API, a bit like cross-platform engines )

[haagch]

I'm not sure how you talked to, but this is all still under heavy discussion internally. No final decisions have been made.

To be perfectly clear: AMD does not know whether what has been dubbed GCN 1.0 and GCN 1.1 will get Vulkan support on Linux?

If they don't, that would be a really sad day. Sure, supporting older hardware is more work, but customers with these GPUs definitely expect to get Vulkan support and many models of this hardware are strong enough to run most current AAA games!

And I have to second the opinion: How can AMD not know that yet? Everyone said that 2015 was targeted for an official release and right now we're already on overtime, waiting for the drivers. Are AMD's Vulkan drivers not yet almost ready?

[artivision]

You could do theoretically do vulkan on pre-GCN parts, but it you wouldn't gain much over existing APIs. Pre-GCN hardware does not have independent compute pipes, so compute and graphics could not be submitted asynchronously and executed in parallel. Shader resources are fixed so you can only have a limited number of resources in flight at any given time and you'd have to explicitly bind the resources at execution time. If someone wants to work on vulkan support for older parts, I certainly won't stop you, just don't expect much of a gain compared to older APIs.

Then all Nvidia GPUs will not manage because they are Tier2 and they don't have Async Compute? Also when I tried some experiments directly on Gallium (not OGL or D3D state trackers), I didn't find out anything that must be fixed.

[agd5f]

Then all Nvidia GPUs will not manage because they are Tier2 and they don't have Async Compute? Also when I tried some experiments directly on Gallium (not OGL or D3D state trackers), I didn't find out anything that must be fixed.

I didn't say it wouldn't work, I said it wouldn't have much advantage over existing APIs on parts with those limitations.

[agd5f]

To be perfectly clear: AMD does not know whether what has been dubbed GCN 1.0 and GCN 1.1 will get Vulkan support on Linux?

If they don't, that would be a really sad day. Sure, supporting older hardware is more work, but customers with these GPUs definitely expect to get Vulkan support and many models of this hardware are strong enough to run most current AAA games!

And I have to second the opinion: How can AMD not know that yet? Everyone said that 2015 was targeted for an official release and right now we're already on overtime, waiting for the drivers. Are AMD's Vulkan drivers not yet almost ready?

Yes, vulkan support is working already. What we have yet to finalize is the exact manner and timelines in which we roll out support for various asic families.

[theriddick]

Well let's hope that AMD is just withholding comment/confirmation pending internal testing of their driver support for older cards.
(can a 390x really be considered OLD????? Its their only cards with 8GB GDDR5 atm!)

[Hibbelharry]

Damn...

I just convinced myself to buy a nice new 27" UHD Display from LG, which got really good reviews. Up to now, I'm using a Radeon 7950 to push pixels to my LCD, working pleasant these days with my 1920x1200 px LCD. I also thought I should maybe get a faster Graphicscard to better suit that new LCD and tried to convince myself to buy a R9 390, but I'm struggling now. If all new stuff will just arrive for GCN1.2, the R9 390 (around 350 Euros for a decent card) is no good idea. I can now opt for tonga, which is quite a bit slower and a bit cheaper, but not absolutely the right card to drive a UHD Display for my occasional gaming needs or buy a absolutely top notch Fiji card (550 Euros) and pay a quite a lot of more money without getting so much more performance. Hell, this is difficult.