https://developer.nvidia.com/reading-between-threads-shader-intrinsics
My point stays. In latest Vulkan drivers for Nvidia hardware there is addition of Shader Intrinsics, and Nvidia GPUs receive boost from it.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
AMD Confirms RX 480 At $199 USD, Other APU & Polaris Announcements
- Thread starter lowendlinux
- Start date
- Sort by reaction score
My point stays. In latest Vulkan drivers for Nvidia hardware there is addition of Shader Intrinsics, and Nvidia GPUs receive boost from it.
I can't tell if you're being intentionally difficult/obtuse here or not. Did you see all the "NV" suffixes/prefixes on all those intrinsics? That means they are NVIDIA specific and unlikely to be supported on any other vendor. If you write GLSL code that uses these NVIDIA extensions, it will not run on an AMD GPU. Similarly, if you use AMD's intrinsics to do similar stuff for GCN, the shaders will not run on NVIDIA (i.e. the driver will refuse to load them because it uses functionality that their GPU doesn't support).
Or, in other words, there are shader intrinsics for all sorts of things, it's not just one big bucket that everyone supports. AMD exposes intrinsics so developers can access the full range of shader instructions for GCN. Similarly, NVIDIA exposes intrindics so developers can access the full range of shader instructions for their GPUs. These are distinct sets of functionality with no overlap at the shading language level, i.e. AMD's intrinsics do not run on NVIDIA and vice versa.
The point here is not that there the vendors provide tools that have been programmed in a hardware-dependent manner, but that if you use AMD's you can port and optimize them yourself for other implementations, and that you can share such development.
Huh? AMD's shader intrisnics will only work on AMD GPUs, and since that's what we've been talking about I think you've got this completely backwards.
For example:
https://www.extremetech.com/gaming/...port-amd-claims-substantial-performance-boost
http://radeon.com/en-us/doom-vulkan/
AMD modified DOOM to use GCN intrinsics and was able to get a substantial performance boost by leveraging these GCN-specific instructions.
To be clear: I am not complaining that AMD went and did this, but I do think it's morally equivalent to NVIDIA providing a set of GameWorks effects that leverage NVIDIA shader intrinsics and thus if you're going to bash NVIDIA for GameWorks perhaps you should rethink your position on AMD as well.
GPUOpen is open source, the community can modify it to adapt it to other hardware implementations.
It looks like it could launch at $549.How come those benchmarks contradict real world usage?
https://wiki.blender.org/index.php/Dev:Source/Render/Cycles/OpenCL
For example here, the link that was debated lately. How come in Final Cut Pro X AMD GPUs are faster while using OpenCL?
How come the benchmarks you use contradict real world usage?
If you will use 11 TFLOPs GPU vs 6TFLOPs, yes, Nvidia GPUs will be faster. But it will be actually interesting to see what will happen when AMD will come up with 11 or higher TFLOPs GPUs.
Why you guys are looking at what was in the past, and not where Apple is going?
And this is not the fastest AMD Card.
P.S. The GPU from AMD in question has maximum of 9.8 TFLOPs. And it trades blows with 11 TFLOPs monster.
Interesting comparison of Video Composition here.
You are correct on the matter that Shader Intrinsics are different for different sort of things.
But not for DX12 and Vulkan. Specs were made not by AMD and MS/Sony and Khronos group with exclusion of Nvidia and Intel. All of IEMs participated in the spec. Funniest part, and biggest clue you have here, that DX12 is based on Mantle is the Shader Intrinsics factor. On both APIs they have exactly the same nature. On all three actually. Vulkan is Mantle with different name. DX12 is DX11 with Mantle and HLSL(soon to be released).
Shader Intrinsics in DX12 and Vulkan have specific nature, to which you optimize differently for different architectures, but the end result is the same for every IEM, but is dependent on performance of the hardware it is optimized for.
There is no denying that there are specific and different versions of Shader Intrinsics for different applications. But DX12 and Vulkan have specific design, and every IEM has to optimize for it.
And biggest, most important part: Optimization of SI for one architecture does not gimp performance for other architectures in DX12 and Vulkan APIs.
Using GameWorks gimps performance on both AMD and Intel GPUs. Only Nvidia is benefiting from it(in the perception of minds, because Project Cars was not able to exceed 60 FPS, in 1080p paired with GTX 980 Ti, so where is performance benefit for end users, for something that does not make games look better in any way, shape or form?). Using Gaming Evolved initiative does not gimp performance of Nvidia and Intel GPUs. It only exploits performance of AMD GPUs. How come? Because GTX 1080 even in Gaming Evolved titles is faster than Fury X. Whereas if we would turn the table - the Fury X would be faster than GTX 1080.
I would love to see that...
Nope. Two reasons. There is no FirePro GPUs being released from AMD right now. All of them are called Radeon Pro. Secondly. Highest end GPU from AMD will top out at 1999$. And it will be dual Vega GPU.Or $5549 for the FirePro version
You seem to not understand what a shader intrinsic function is, so let me try and explain again using DOOM as an example. Intrinsics are special low-level hardware instructions or features that are not exposed through an API like DirectX or Vulkan. The vendor (AMD in this case) provides an extension that exposes these intrinsics, so software developers can have full access to the hardware, just like they do on consoles. Let's pretend that AMD has an intrinsic called "Make DOOM run magically faster" that is specific to GCN, and that no other hardware vendor has. AMD then goes and modifies the DOOM engine to use this intrinsic, and as you can see, they made it run roughly 20 to 30% faster:
Optimizing this application using AMD intrinsics absolutely does "gimp performance" for the other vendors (Intel and NVIDIA) because they added a special code path to make it run faster on AMD only. Since this intrinsic is only found in GCN, Intel and NVIDIA can't just "optimize for it" because it's a hidden low-level (probably patented) feature that they don't have. That link even shows that NVIDIA did not get any faster with Vulkan because it was already GPU limited like their OpenGL version was, but they weren't able to use the magic AMD intrinsic path to get a perf boost.
There are other classes of optimizations that both AMD and NVIDIA helps game developers with that provide increased performance on all vendors. However, shader intrinsics are not one of them, because they are by definition specific to one vendor.
This will be my last attempt to try and explain all this stuff, I'm assuming you'll just ignore it and keep pretending that AMD is wonderful and NVIDIA are evil as before.
Why do you read where I write: "Shader Intrinsics do not gimp performance on any of vendors because they are cross platform" as the possibility that Nvidia would use AMD path?You seem to not understand what a shader intrinsic function is, so let me try and explain again using DOOM as an example. Intrinsics are special low-level hardware instructions or features that are not exposed through an API like DirectX or Vulkan. The vendor (AMD in this case) provides an extension that exposes these intrinsics, so software developers can have full access to the hardware, just like they do on consoles. Let's pretend that AMD has an intrinsic called "Make DOOM run magically faster" that is specific to GCN, and that no other hardware vendor has. AMD then goes and modifies the DOOM engine to use this intrinsic, and as you can see, they made it run roughly 20 to 30% faster:
Optimizing this application using AMD intrinsics absolutely does "gimp performance" for the other vendors (Intel and NVIDIA) because they added a special code path to make it run faster on AMD only. Since this intrinsic is only found in GCN, Intel and NVIDIA can't just "optimize for it" because it's a hidden low-level (probably patented) feature that they don't have. That link even shows that NVIDIA did not get any faster with Vulkan because it was already GPU limited like their OpenGL version was, but they weren't able to use the magic AMD intrinsic path to get a perf boost.
There are other classes of optimizations that both AMD and NVIDIA helps game developers with that provide increased performance on all vendors. However, shader intrinsics are not one of them, because they are by definition specific to one vendor.
This will be my last attempt to try and explain all this stuff, I'm assuming you'll just ignore it and keep pretending that AMD is wonderful and NVIDIA are evil as before.
Shader Intrinsics are part of the engine. Part of the API. Each of vendors do optimize for their own, individual paths. But the idea is "unified". Optimizing Shader Intrinsics does not gimp performance for one vendor by exposing others or the other way around. It is just stupid to believe that is the case! Shader Intrinsics are there to expose otherwise invisible capabilities of the GPUs. But the bypassing of DXIL(or any other API) is universal, because it is part of the nature of the API's.
I have made a mistake. I thought that Intrinsics are not available to DX11, but they are. They are not available for OpenGL. Thats why you see boost in performance in Vulkan on AMD hardware, and nothing on Nvidia.
Edit. This is the point I was making previously about the difference of approaches between Nvidia and the rest of industry. Games using Gaming Evolved initiative are designed to work on different paths for each of IEMs involved in optimization of the game engine. Games using GameWorks are designed to work on Nvidia paths, regardless of IEM's involved in optimization of the game engine.
That is exactly what is crippling performance in GameWorks titles, and what is not crippling performance in other titles.
Last edited:
So let me get this straight, you're saying:
- Intrinsics are part of the standard API.
- AMD changing DOOM to use their GCN intrinsics helps all vendors because it's part of the "Gaming Evolved" initiative.
- NVIDIA is bad because GameWorks.
Edit: And to be clear, I don't see anything wrong with either approach, obviously each vendor is going to want to make applications run as fast as possible on their GPUs. I just think it's silly to claim AMD is pure and good while NVIDIA is evil and bad because they take different approaches to achieve the same goal.
AMD did not changed anything.
There is nothing in them that "change" execution for other vendors. It changes only how the code is executed for specific architecture. The same thing is happening for Nvidia and Intel. And nothing here interferes with other IEM's.
Shader Intrinsics are not part of Gaming Evolved initiative. Gaming Evolved was used as an example of not gimping performance for other vendors, contrary to GameWorks.
The biggest problem I think you guys have is with understanding that I am not saying anywhere that something is good or bad. Its you who is making it out of my words. Where I am just stating facts, that GameWorks gimps performance for other vendors, while Gaming Evolved, and Shader Intrinsics in Vulkan/DX11/DX12 do not. You claim that I do not understand what are Shader Intrinsics, and it appears that you are not able to understand broader point of view, and actually understand what are Shader Intrinsics, and how they are implemented in the Game Engine/API's.
So yes, lets agree to disagree.
So according to that video, the RX 580 is < 2% faster across all the games tested and requires 60 W more than the gtx 1060. One of those cards would fit in an iMac, the other wouldn't.
In other bad news for AMD, the new game Prey which AMD is using to market its cards (and the upcoming Vega) runs significantly better on Nvidia.
So according to that video, the RX 580 is < 2% faster across all the games tested and requires 60 W more than the gtx 1060. One of those cards would fit in an iMac, the other wouldn't.
In other bad news for AMD, the new game Prey which AMD is using to market its cards (and the upcoming Vega) runs significantly better on Nvidia.
And to be clear, that 60W is 50% of the 1060's power consumption (120W vs 180W). I really don't understand how people can claim AMD has a more efficient architecture.
I am very interested to see how Vega shapes up. This is basically AMD's first architecture really designed for the 14/16 nm generation. Polaris 10/11 is more of a stop gap solution. Vega basically has to bring AMD back to relative parity with Nvidia, otherwise they are in big trouble.
It appears you forgot compute factor.
Not to mention that RX 480 was possible to find aftermarket at 160$. 10$ more, than GTX 1050 Ti, over which RX 480 runs circles around both in gaming and compute.
You're the one who posted a video with a bunch of game benchmarks, that once again showed the 480 losing to the 1060 on average across all games. Raw TFLOPs doesn't really matter that much in many applications, such as all these games that were tested.
Also, think about why they had to lower the price by that much. How profitable was AMD last quarter again? Oh that's right, they're still losing money hand over fist.
Seriously, there's 1% difference on average, with about half the games going each way.
Confirmation Bias.
Funiest part is elsewhere.
RX 480 is loosing vs GTX 1060 in games pre 2016. In most games 2016+ it is winning with GTX 1060, with exclusion of GameWorks titles, and Nvidia optimized titles.
How can this be?!
People are looking at simple 1% difference and will say: RX 480 is rubbish it loses to GTX1060. And then you will look into the games tested, and ask yourself a question: what methodology was used to draw conclusions?
I am mostly staggered by one thing. Overwatch was gigantic win for Nvidia in last year. RX 480 was 40 FPS slower than GTX 1060 in Ultra Settings 1080p, and 15-20 FPS in Epic settings.
Currently the GPUs are on par in this game, in Epic settings.
I have brought this benchmark suite to prove you guys two things. Software matures with time, and AMD GPUs are becoming faster with time, and are outpacing Nvidia GPUs, at competitive price brackets, with time. It is historically proven fact, right now. It will not be long when we will see that RX 480 is faster across the board than GTX 1060.
Im wondering what people will then say about it?
P.S. Everything in big picture is linked directly to what we have been discussing over last few pages of this thread. You just have to open your eyes guys.
Last edited:
You are making a lot of assumptions about "we" are thinking. I didn't say the RX 480 was rubbish, I said that that the GTX 1060 is much more efficient. Given that Apple tends to design thermally constrained systems, this matters a great deal.