Are the next gen consoles the realization of AMDs HSA dream?

[BD2003]

Been thinking about this a bit, hoping the more technically inclined here can clarify a few things.

So my understanding of HSA is that it begins with tightly integrating the CPU and a compute capable GPU onto the same die, and writing to the same memory space. This would then enable programs to take advantage of whichever computational units are best for the job without the losses associated with copying back and forth from a CPU to a discrete GPU. In same cases this is supposedly a huge boost in efficiency, particularly physics. So....

Is there merit to the idea to begin with? Does an APU really have potential to be much faster in some cases than just a CPU or CPU + dGPU?

If so, because the consoles are a fixed platform, there's no need to ask the question of whether or not devs will take to coding to take advantage of it....they will. Does the architecture in the upcoming consoles (jaguar + GCN) require them to hand code specific to this capability? Or does it happen fairly automatically at the compiler level, or completely automatically at the chip level?

If hand coding, doesn't that make AMD kind of the de facto leader for gaming? These optimizations are going to happen....how will that filter down to the PC? Does it require API support in DirectX? Can that code even run on Intel/Nvidia? Or do we get to the point where we start to see APU + dGPU working in tandem, instead of the iGPU essentially dead weight in the presence of a dGPU as it is now?

It always seemed to me like they were on to something big....but I'm not a programmer. Getting in both new consoles looks like a validation of it all to me. Could we potentially be seeing a resurgence of AMD, perhaps not in efficiency, but at least in gaming performance? Not flamebait, genuinely curious.

 

[Olikan]

at least at PS4, there are some important HSA features missing...
like graphics pre-emption and context switch

but...don't worry that much, desktops will brute force it's way, they were doing this since the beggining

even if latency became impossible to brute force, nothing stops amd or nvidia put some small cpu cores inside theyr gpu cards and emulate a console

 

[BrightCandle]

Is there merit to the idea to begin with? Does an APU really have potential to be much faster in some cases than just a CPU or CPU + dGPU?

In theory for certain types of programs now it should in theory be quicker. In practice of course the CPU is always going to be sharing thermal headroom with the GPU for the forseeable future and hence a discrete card will ave more computation resources on offer but with higher initial latency. Thus small programs which use limited amounts of computation will be faster today, marginally.

Does the architecture in the upcoming consoles (jaguar + GCN) require them to hand code specific to this capability?

Yes it does, and its not easy to code for. It requires a complete rewrite of any code you port to the approach. Its also pretty limited in what it can do efficiently so you have to be careful about how the programs are written as well.

Or does it happen fairly automatically at the compiler level, or completely automatically at the chip level?

No definitely no.

If hand coding, doesn't that make AMD kind of the de facto leader for gaming?

Seeing as how NVidia currently holds most of the compute market I would call them the actual leader. AMD might have managed to sell low end hardware to the console makers but they are far from leading, they are just cheaper.

These optimizations are going to happen....how will that filter down to the PC?

Seeing as how the PC has had this capability for many years and some games already use it this is kind of an odd question. Its already there

Does it require API support in DirectX?

Already there.

Can that code even run on Intel/Nvidia?

Nvidia got there first. Intel has their own discrete card especially for this but X86 and also openCL works on Intel's latest CPUs. Plenty of ways to cut this problem but in essence they all have the capability to do this, its nothing new.

Or do we get to the point where we start to see APU + dGPU working in tandem, instead of the iGPU essentially dead weight in the presence of a dGPU as it is now?

Unlikely because they all speak very different things. Going to require a lot of future work to make the cards cooperate on the same workloads and I don't see that happening anytime soon.

Could we potentially be seeing a resurgence of AMD, perhaps not in efficiency, but at least in gaming performance?

AMD never stopped doing well with its graphics performance, it has been and remains competitive there. It beats Intel and competes on performance and price well with NVidia. They still have issues with quality control both on the hardware and especially in their software but they competitive and selling well.

 

[Arachnotronic]

AMD never stopped doing well with its graphics performance, it has been and remains competitive there. It beats Intel and competes on performance and price well with NVidia. They still have issues with quality control both on the hardware and especially in their software but they competitive and selling well.

*giggle*

You don't read AMD's financial statements, do you? Nvidia is slaughtering AMD in GPU sales. That being said, AMD's hardware is fine, but Nvidia's got the brand equity under lock & key.

 

[BD2003]

So assuming the game devs do go through the trouble of hand coding this, which I'm sure they will over time, given the fixed platform - the techniques they will come up with will apply equally as well to CPU + dGPU, but potentially not as efficiently due to the split? Then its not really a matter of AMD going off in a separate direction from intel/nvidia, just applying a different method to get to the same place?

 

[metrix]

To start out: I am a programmer, and I look forward to writing software for Kaveri!

In theory for certain types of programs now it should in theory be quicker. In practice of course the CPU is always going to be sharing thermal headroom with the GPU for the forseeable future and hence a discrete card will ave more computation resources on offer but with higher initial latency. Thus small programs which use limited amounts of computation will be faster today, marginally.

You are assuming that a CPU and a GPU handle the same type of workload. When Kaveri is released, a programmer will be able to write software that is able to utilize the HSA infrastructure and get massive performance boosts on parallel-able pieces of a coding project without the need to totally re-write the software.

Yes it does, and its not easy to code for. It requires a complete rewrite of any code you port to the approach. Its also pretty limited in what it can do efficiently so you have to be careful about how the programs are written as well.

Take a look at this: http://techreport.com/news/23673/amd-teams-with-oracle-on-java-acceleration-qualcomm-on-hsa

Most applications will have to be modified (i'm not sure how much) but they won't require a complete re-write to get enhanced performance out of an APU.

Originally Posted by BD2003
Or does it happen fairly automatically at the compiler level, or completely automatically at the chip level?

No definitely no.

I have read that most of the changes will be dealt with within the Programming language, and that programmers will just need to "flag" which methods would work better on the GPU versus the CPU.

Originally Posted by BD2003
These optimizations are going to happen....how will that filter down to the PC?

Seeing as how the PC has had this capability for many years and some games already use it this is kind of an odd question. Its already there

The PC has had graphics cards for years, but PC's havn't had a graphics card SHARE memory with a processor, and the ability to program this type of thing hasn't been built directly into mainstream programming languages.

To run a parallel data set on a computer right now would look something like this:
Main memory -> PCI BUS(latency) -> GPU computes task -> PCI BUS(latency) -> main memory

where this same task on Kaveri would run like:
Main memory -> GPU computes task -> Main memory

it doesn't seem like much to copy data across the bus, but most applications will have to do that thousands of times a second.. The latency adds up. Tie this in with the current difficulty of coding software for a GPU and you can see that AMD is sitting on something big! Imagine the performance boost and the added battery life of an Android tablet device running on a HSA enhanced Java JVM. :biggrin:

 

[beginner99]

To start out: I am a programmer, and I look forward to writing software for Kaveri!


You are assuming that a CPU and a GPU handle the same type of workload. When Kaveri is released, a programmer will be able to write software that is able to utilize the HSA infrastructure and get massive performance boosts on parallel-able pieces of a coding project without the need to totally re-write the software.

This is however only relevant for low-level languages like C/C++. The actual benefits will come when .net and Java runtime make use of this stuff. And considering the track record there that could be a while.

I doubt HSA will be used much in PS4. Because if you run physics on the gpu you will end up with a worse image or lower FPS due to less gpu resources for actual graphics processing. (Unless the used jaguar cores actually keep their gpu part, then it would make sense).

But this is the general fallacy of APUs. If you start using the gpu for other stuff you will have unpredictable graphics performance....and the real-world problems APUs solve can be solved with dedicated hardware (speak QuickSync) or new cpu uArch better suited for such loads (haswell).

 

[mrle]

This is however only relevant for low-level languages like C/C++. The actual benefits will come when .net and Java runtime make use of this stuff. And considering the track record there that could be a while.

Actually, things are moving forward there too, for example there is already a quite usable OpenCL Java API from AMD : http://developer.amd.com/tools/heterogeneous-computing/aparapi/

 

[grimpr]

This is however only relevant for low-level languages like C/C++. The actual benefits will come when .net and Java runtime make use of this stuff. And considering the track record there that could be a while.

I doubt HSA will be used much in PS4. Because if you run physics on the gpu you will end up with a worse image or lower FPS due to less gpu resources for actual graphics processing. (Unless the used jaguar cores actually keep their gpu part, then it would make sense).

But this is the general fallacy of APUs. If you start using the gpu for other stuff you will have unpredictable graphics performance....and the real-world problems APUs solve can be solved with dedicated hardware (speak QuickSync) or new cpu uArch better suited for such loads (haswell).

Nope, the GPU in Kaveri will have pre-emption and will run smoothly graphics and compute workloads without pauses,glitches and huge latencies.

 

[NTMBK]

Seeing as how NVidia currently holds most of the compute market I would call them the actual leader. AMD might have managed to sell low end hardware to the console makers but they are far from leading, they are just cheaper.

The compute market has nothing to do with gaming leadership...

 

[NTMBK]

This is however only relevant for low-level languages like C/C++. The actual benefits will come when .net and Java runtime make use of this stuff. And considering the track record there that could be a while.

I doubt HSA will be used much in PS4. Because if you run physics on the gpu you will end up with a worse image or lower FPS due to less gpu resources for actual graphics processing. (Unless the used jaguar cores actually keep their gpu part, then it would make sense).

But this is the general fallacy of APUs. If you start using the gpu for other stuff you will have unpredictable graphics performance....and the real-world problems APUs solve can be solved with dedicated hardware (speak QuickSync) or new cpu uArch better suited for such loads (haswell).

There are parts of the graphics pipeline currently done on CPU, despite the fact that they are parallel algorithms which would run far better on a GPU- because the cost of shuffling data backwards and forwards would kill performance. That cost is gone in an APU. Take a look at these slides for some more information (look at the sections on frustum culling, occlusion culling, asset decompression): http://www.slideshare.net/zlatan4177/gpgpu-algorithms-in-games

 

[Abwx]

Actualy it s already a fully compatible HSA item.
On another note , the chip must be huge.

“The PlayStation 4 blew me away but I am conscious about being too positive about it because of course we have other partners that are competing. […] It looks like they did an extremely good job of the engineering. The CPU and the GPU are on the same chip, which solves a lot of the problems, e.g., you do not have to toss things over to a graphics card and try to get them back and forth,” said Mr. Gustafson in an interview with Canada.com web-site.

Sony PlayStation 4 is based on a semi-custom AMD Fusion system-on-chip that integrates eight AMD x86 Jaguar cores, custom AMD Radeon HD core with unified array of 18 AMD GCN-like compute units

http://www.xbitlabs.com/news/multim...ism_Thanks_to_Heterogeneous_Architecture.html

 

[aaksheytalwar]

What is HSA?

 

[Idontcare]

What is HSA?

A dream that fits nicely inside a pipe, for now, for which a multitude of businesses are keeping their finger in the pot just in case it turns out to be the "next thing". For now, it has been the future (well, that and fusion) for a good 7 yrs thus far.

Its a volume business you see But someday someone might make money on it. So far only ATI's ex-shareholders have made money on it.

 

[R0H1T]

So far only ATI's ex-shareholders have made money on it.

Where's Hector these days :sneaky:

 

[inf64]

We have hit the IPC wall, soon we will hit the process node wall ( and "moar cores" will become more difficult to bolt on a die within reasonable TDPs) . The only thing left is coprocessor model with specialized fixed function HW and GPGPU. HSA is just that, heterogeneous model to harness the fp power of GPUs that share the same memory address space for most efficient computing on GPU. Programming is one difficulty which is being worked on now (worked hard by HSA consortium).

 

[AnandThenMan]

A dream that fits nicely inside a pipe, for now, for which a multitude of businesses are keeping their finger in the pot just in case it turns out to be the "next thing". For now, it has been the future (well, that and fusion) for a good 7 yrs thus far.

Its a volume business you see But someday someone might make money on it. So far only ATI's ex-shareholders have made money on it.

I'm not sure if you are being serious or not. But I'll go with you trying for extreme sarcasm, although I can't imagine exactly why.

HSA is very very exciting and is the future of computing.

 

[Dribble]

The PS4 easily seems the closest AMD have got to "Fusion" yet - the hardware seems to be there and because every PS4 has that AMD chip in the devs will code to take advantage of all the AMD specific features (no need to code to suit Intel here). Seems like it has even more potential then the cell (which in the end offered nothing that the more conventional xbox 360 cpu couldn't do).

 

[Abwx]

They said that they may sell cut down version of the ¨PS4 chip ,
surely to recycle deffective dies , so we can well end with 6 cores
and 800 or so GCN SPs for Kabini boards if ever they keep on this project.

 

[MisterMac]

The PS4 easily seems the closest AMD have got to "Fusion" yet - the hardware seems to be there and because every PS4 has that AMD chip in the devs will code to take advantage of all the AMD specific features (no need to code to suit Intel here). Seems like it has even more potential then the cell (which in the end offered nothing that the more conventional xbox 360 cpu couldn't do).

Why do AMD hands keep heralding HSA as amazing - and think that weird shit like game\application programmers program for Intel only?

Are you that blind?


Is AMD's SSE hardware implementation so superbly inferior to Intel's implementation of SSE because Intel spec'ed it?


Jesus christ.

If HSA actually gets a standard(it has none) - EVERYONE will use it.
Including Intel & nVidia.

You cannot patent the unified shader model or HSA.
And if you want traction you need a agreed standard between all parties involved - and that means all parties will use the same grounds.

 

[Idontcare]

HSA is very very exciting and is the future of computing.

I know, I've been told so, frequently, since 2006. Can't wait for the future to get here.

 

[AnandThenMan]

I know, I've been told so, frequently, since 2006. Can't wait for the future to get here.

Innovation doesn't always occur as quickly as we'd like. And industry forces often delay much needed innovation from reaching a critical mass. At one time direct current was being touted as the best solution for the power grid.

 

[ShintaiDK]

Innovation doesn't always occur as quickly as we'd like. And industry forces often delay much needed innovation from reaching a critical mass. At one time direct current was being touted as the best solution for the power grid.

Really? I thought Tesla won that flat out, since we started with DC, but the localization of the power plants outright killed it.

But as IDC says, HSA is still a pipe dream and it continues to be so. We are also still waiting on BitBoys.

 

[Enigmoid]

Nope, the GPU in Kaveri will have pre-emption and will run smoothly graphics and compute workloads without pauses,glitches and huge latencies.

Possibly but thats not answering the question. If you try to throw some physics at the gpu that will take 20% of the gpu's power, frame rate will drop 20%.

 

[AnandThenMan]

Really?

Yes.

During the initial years of electricity distribution, Edison's direct current was the standard for the United States, and Edison did not want to lose all his patent royalties.

Edison carried out a campaign to discourage the use[26] of alternating current, including spreading disinformation on fatal AC accidents, publicly killing animals, and lobbying against the use of AC in state legislatures. Edison directed his technicians, primarily Arthur Kennelly and Harold P. Brown,[27] to preside over several AC-driven killings of animals, primarily stray cats and dogs but also unwanted cattle and horses. [28] Acting on these directives, they were to demonstrate to the press that alternating current was more dangerous than Edison's system of direct current.[29] He also tried to popularize the term for being electrocuted as being "Westinghoused". Years after DC had lost the "war of the currents," in 1903, his film crew made a movie of the electrocution with high voltage AC, supervised by Edison employees, of Topsy, a Coney Island circus elephant which had recently killed three men.

Edison opposed capital punishment, but his desire to disparage the use of alternating current led to the invention of the electric chair. Harold P. Brown, who was being secretly paid by Edison, built the first electric chair for the state of New York to promote the idea that alternating current was deadlier than DC.

read more​

But as IDC says, HSA is still a pipe dream and it continues to be so.

No it's happening right in front of your eyes, you just need to open them.