Maybe We Don't Understand the implications of OCL and HSA

[NTMBK]

By that logic everything now should support 6+ threads. Because Xbox360 and PS3 did.

HSA has gone nowhere. OpenCL is just a copy of CUDA in terms of apps.

So what will radically change?

What the PS3 had was nothing like 6+ general purpose threads- it had only two general purpose threads (one of which was an SMT thread, so much less useful), and the Cell's SPUs which were only useful for specific tasks- in terms of how you'd structure your engine (offloading tasks from the main core to accelerators), it was actually closer to the GPGPU model. The two general purpose threads of the Cell were the lowest common denominator, which is why we didn't see many engines optimised that well for more than that.

 

[SammichPG]

By that logic everything now should support 6+ threads. Because Xbox360 and PS3 did.

HSA has gone nowhere. OpenCL is just a copy of CUDA in terms of apps.

So what will radically change?

Good multithreading takes a lot of work, HSA should be a matter of flipping a switch in the compiler/middleware of your game engine and forget about it.

Every game engine running on consoles will have HSA and their windows counterpart should have it as well.

 

[NTMBK]

Good multithreading takes a lot of work, HSA should be a matter of flipping a switch in the compiler/middleware of your game engine and forget about it.

I wouldn't believe that hype just yet. Wait until we actually see a working HSA compiler, and then start celebrating how easy it is to use... We've seen plenty of promises like this before (like how auto-vectorization will make all our software go 4 times faster).

 

[SammichPG]

I wouldn't believe that hype just yet. Wait until we actually see a working HSA compiler, and then start celebrating how easy it is to use... We've seen plenty of promises like this before (like how auto-vectorization will make all our software go 4 times faster).

I agree, just stating what amd has promised (heh).

We can be sure that at least consoles will use HSA, we'll see how useful it's going to be for the pc/server markets.
Imo HSA is amd's last chance at not ending like via or one of the smaller arm chipmakers with their gpu business bought from someone like qualcolmm.

AMD failure would be a disaster for intel, the consumer choice would be between their market segmentation and ARM socs that are getting closer and closer to be fast enough for desktop usage (around core 2 duo level or even less).

Add cheap 64bit ARM chips with features needed in the server world and intel wouldn't be able to keep its process lead without the volume coming from millions of pentiums and celerons.

What's intel without process tech lead? A huge giant battling with nimble opponents who are happy with lower margins... it reminds me of Intel vs DEC vs Sun.

People do most of their computing on stuff that's slower than a atom and they pay stupid prices for the convenience of a compact device, it's not about power anymore.

 

[Cerb]

Good multithreading takes a lot of work, HSA should be a matter of flipping a switch in the compiler/middleware of your game engine and forget about it.

:biggrin:

No. To just flip a switch, we would need the CPU and GPU to share the same thread. HSA could bring it from, "good CPU and GPU multithreading is damn near impossible with shared memory," down to, "good multithreading takes a lot of work."

But, by a compiler setting? That's basically AVX2 (along with plenty of profiling, and assuming mostly new code, rather than trying to make old code run faster).

 

[mrmt]

Did OpenCL slow down that IB? Else it makes no sense.

AMD marketing at work again.

http://www.adobe.com/products/premiere/tech-specs.html

No Intel GPU is supported in Adobe Premiere Pro CC, so IVB is running all OpenCL code on the CPU. Now what's left is AMD trick to make an anemic APU goes on par with an Intel quad core.

 

[ShintaiDK]

AMD marketing at work again.

http://www.adobe.com/products/premiere/tech-specs.html

No Intel GPU is supported in Adobe Premiere Pro CC, so IVB is running all OpenCL code on the CPU. Now what's left is AMD trick to make an anemic APU goes on par with an Intel quad core.

I wonder what the performance with OpenCL off for the IB would have been. I am sure there is a reason why AMDs marketing didnt include it.

 

[sontin]

AMD marketing at work again.

http://www.adobe.com/products/premiere/tech-specs.html

No Intel GPU is supported in Adobe Premiere Pro CC, so IVB is running all OpenCL code on the CPU. Now what's left is AMD trick to make an anemic APU goes on par with an Intel quad core.

What do you expect? AMD did it with Winzip 16.5 and OpenCL, too. It was only supported on their plattform...

The Roy Taylor guy said that "open standards benefit all" yet it's AMD who is using open standards to sabotage competition...

 

[A5]

It isn't AMD's job to get that stuff working on Intel's chips. If Intel wants people to use OpenCL on their iGPUs, they can help developers learn how to do it or provide the tools to make it easier.

 

[mrmt]

It isn't AMD's job to get that stuff working on Intel's chips. If Intel wants people to use OpenCL on their iGPUs, they can help developers learn how to do it or provide the tools to make it easier.

Kudos to AMD for spending time and money with developers to optimize for their platform. Nobody is questioning AMD reasoning here. But another very different thing is to misrepresent benchmark data, and this is what they are doing with that slide. They could just have ommited the OpenCL part on Intel benchmark data or just mentioned that Premiere doesn't use Intel IGP for OpenCL, but instead they put OpenCL on both data as if they were the same thing. This is misleading marketing at work again, but that's AMD, we cannot expect them to be much more honest, can we?

I also found rather strange that without OpenCL the 6800k is on par with a 3470. Can someone run the benchmarks to see if the results are really comparable?

 

[mikk]

Basically Cyberlink is being lazy here. When they use AMD's video encode API they call it OpenCL, when they use NVIDIA's they call it CUDA, and when they use Intel's they call it QuickSync. In practice with current drivers ME supports NVENC, VCE, and QuickSync.

Do you have proof? I don't think MediaEspresso support NVENC and VCE. Of course they shipped a test version to reviewers long time ago but in public versions as far as I know it isn't supported. In the current MediaEspresso version on a Kepler GPU hardware encoding is pretty slow compared to Quicksync.

 

[USER8000]

What do you expect? AMD did it with Winzip 16.5 and OpenCL, too. It was only supported on their plattform...

The Roy Taylor guy said that "open standards benefit all" yet it's AMD who is using open standards to sabotage competition...

AMD did not invent OpenCL,and it is an open standard maintained by the Khronos Group which was founded in 2000 by many companies including Intel and ATI. Apple originally started development of OpenCL and has pushed for its use for the OS X version of the Adobe CS and this appeared a while back,so this is an extension of this work which AMD facilitated for the Windows version.

Moreover,OpenCL does run on the newer Intel IGPs last time I checked,so ultimately any performance defiencies of the IB IGP using OpenCL are down to Intel in the end,and not AMD. If Intel improves OpenCL performance of its IGPs it will be reflected in the benchmarks anyway.

 

[mikk]

AMD did not invent OpenCL

Key point is optimization. Handbrake QS last year was made by AMD. No surprise that on Intel OpenCL didn't work at that time. Here the same, AMD is directly involved. Because of this you can't expect that OpenCL on Intel (or Nvidia) is working great (if it works). As long as AMD is directly involved other vendors can't perform good or they are not supported.

AMD and Adobe are working together to help provide powerful solutions that enable creative professionals to speed workflows and increase throughput

 

[USER8000]

Key point is optimization. Handbrake QS last year was made by AMD. No surprise that on Intel OpenCL didn't work at that time. Here the same, AMD is directly involved. Because of this you can't expect that OpenCL on Intel (or Nvidia) is working great (if it works). As long as AMD is directly involved other vendors can't perform good or they are not supported.

OpenCL performance is stellar with the new Haswell chips.

http://www.tomshardware.com/reviews/core-i7-4770k-haswell-review,3521-3.html

Dunno why you have to bash it so badly. You'll get OCL in virtually any device from now on and HSA in almost any non Intel/Nvidia stuff. As far as I got it, HSA is only a way to speed up even further OpenCL and the likes.

Look at the link posted earlier.

Its only a three page thread too,LOL and the conspiracies theories are out already.

Also,Intel,AMD or Nvidia marketing not trying to make their products look the best,is happening as much as the likelihood of seeing supersonic flying pigs powered by fusion engines.

Hence as everyone I know does,you take it for a pinch of salt until reviews come out to see whether the truth is anywhere near the reality.

However,it does appear Intel also like AMD and Nvidia,are taking GPU compute more seriously,so as a trend it is going to make more and more of an impact on everyday applications I suspect. Even the HSA Foundation does have a number of companies on the bandwagon including those like Qualcomm who could be counted as competitors to AMD and Nvidia too.

 

[mikk]

Look at the link posted earlier.

Yes there is a confirmation. In general or in particular Haswell is doing great compared Trinity when it comes to OpenCL as long as there is a fair implementation without a big vendor intervening. The new collaboration with Adobe and AMD refer to upcoming Adobe products. We should expect Intel doing much worse there.

 

[USER8000]

Yes there is a confirmation. In general or in particular Haswell is doing great compared Trinity when it comes to OpenCL as long as there is a fair implementation without a big vendor intervening. The new collaboration with Adobe and AMD refer to upcoming Adobe products. We should expect Intel doing much worse there.

There would probably be no OpenCL accelerated Windows version of the Adobe CS in the first place if AMD was not working with Adobe. What you don't realise is that this work has been going on for a while and there were demoes of AMD GPUs running it with OpenCL activated ages ago.

Here is an example in 2012:

http://www.youtube.com/watch?v=J6MroxhsLT0

That is Adobe demoeing parts of it in October 2012 and talking about their collaboration with AMD using OpenCL.

Despite this,there is no evidence from reviews to show any artificial reduction in performance of Intel IGPs using OpenCL with Adobe products. If anything the improvements in the HD4600 show how relatively poor the HD4000 and HD3000 were in this regard,and indicates Intel knew this,improved performance of the Haswell IGPs in this regard by a massive amount and hence I am not going to argue with them.

There is NO reason for Adobe to cap performance on Intel IGPs,especially since many users are running the CS and CC on Intel boxes including laptops,many of which use IGPs,specifically Intel IGPs. Even looking at the link,it seems Winzip is the same too.

Of course since this is the internet everything is some conspiracy,rather than the much more boring reality.

 

[ViRGE]

We can be sure that at least consoles will use HSA, we'll see how useful it's going to be for the pc/server markets.
Imo HSA is amd's last chance at not ending like via or one of the smaller arm chipmakers with their gpu business bought from someone like qualcolmm.

Actually they won't. The console APUs are not HSA capable. First generation Jaguar is not HSA ready, among other limitations.

The first HSA capable APU is Kaveri, and even that's going to be a bit of a stretch since it's more like HSA 0.9 (the GPU won't be capable of fast context switching).

 

[Khato]

The first HSA capable APU is Kaveri, and even that's going to be a bit of a stretch since it's more like HSA 0.9 (the GPU won't be capable of fast context switching).

Oh? Hadn't heard that tidbit before. Wonder when they'll have that working properly...

 

[Imouto]

It's so fun how people get away in this forum making such claims without a single link supporting them.

 

[Hitman928]

Actually they won't. The console APUs are not HSA capable. First generation Jaguar is not HSA ready, among other limitations.

The first HSA capable APU is Kaveri, and even that's going to be a bit of a stretch since it's more like HSA 0.9 (the GPU won't be capable of fast context switching).

Are you saying, then, that the 8GB of RAM won't actually be unified? I know that AMD/Sony/MS haven't really announced anything hsa related, but most people (tech reporters) seem to think that the consoles will have support for hsa. These are semi-custom chips which are coming out about half a year after jaguar. I'm not saying you're wrong, I just haven't seen anything to lean me one way or the other yet.

 

[monstercameron]

Are you saying, then, that the 8GB of RAM won't actually be unified? I know that AMD/Sony/MS haven't really announced anything hsa related, but most people (tech reporters) seem to think that the consoles will have support for hsa. These are semi-custom chips which are coming out about half a year after jaguar. I'm not saying you're wrong, I just haven't seen anything to lean me one way or the other yet.

hsa isnt just one thing a but whole host of concepts, as virge speculate the console apu might not have "fast context switching," he didnt say it didnt have huma...

aside: maybe its too early to have this discussion, this thread isnt going anywhere...

 

[Phynaz]

Are you saying, then, that the 8GB of RAM won't actually be unified?

Don't miss the difference between unified memory and unified address space. They aren't the same thing.

http://developer.amd.com/wordpress/media/2012/10/hsa10.pdf

 

[Hitman928]

Don't miss the difference between unified memory and unified address space. They aren't the same thing.

http://developer.amd.com/wordpress/media/2012/10/hsa10.pdf

I understand that, I just didn't clarify what I meant by 'unified' That was my mistake.

 

[galego]

Actually they won't. The console APUs are not HSA capable. First generation Jaguar is not HSA ready, among other limitations.

The first HSA capable APU is Kaveri, and even that's going to be a bit of a stretch since it's more like HSA 0.9 (the GPU won't be capable of fast context switching).

In both consoles, the CPU and GPU will be on the same die (an AMD APU). Just as the PS4 has 8GB of high-speed memory that is shared by the CPU and GPU, the Xbox One, by virtue of being based on the same APU heterogeneous system architecture (HSA), will probably be the same.

http://www.extremetech.com/gaming/156273-xbox-720-vs-ps4-vs-pc-how-the-hardware-specs-compare

 

[ViRGE]

http://www.extremetech.com/gaming/156273-xbox-720-vs-ps4-vs-pc-how-the-hardware-specs-compare

http://images.anandtech.com/doci/5491/Screen Shot 2012-02-01 at 2.14.16 PM.png

http://images.anandtech.com/doci/5503/Screen Shot 2012-02-02 at 3.12.58 PM.png

http://images.anandtech.com/doci/5493/Screen Shot 2012-02-02 at 9.20.35 AM.png

The HSAIL isn't even finalized yet; version 0.95 was released just last month. Kaveri is the first APU that is HSA capable; first gen Jaguar products are not going to be HSA capable. Sites like ExtremeTech are essentially failing to understand what HSA is; unified memory is a part of HSA. It is not all of HSA.