News - Sony to dump NVIDIA for AMD as GPU chip provider for PS4?

[HEXUS]

AMD apparently working with Sony on a "hush-hush" next-generation console.

Read more.

[Scribe]

So the real wonder on my mind is, will the PS4 be receiving some graphical love only or will be be seeing something like a top-of-the-range or bespoke Trinity APU in the system.

Would certainly keep costs down and make life easier for the developer but would it provide enough raw horsepower?

We could end up seeing some Kaveri Steamrollers with some HSA fun if it's only in the early developmental stages: http://hexus.net/tech/news/industry/34769-amds-2012-2013-roadmap-2013-28nm/

[abaxas]

Surely this will be the round of commodity consoles?

Cheap to make, cheap to develop for but powerful enough for all but the richest developer.

[Brewster0101]

So the real wonder on my mind is, will the PS4 be receiving some graphical love only or will be be seeing something like a top-of-the-range or bespoke Trinity APU in the system.

Would certainly keep costs down and make life easier for the developer but would it provide enough raw horsepower?

How about 2 or maybe 4 quad core Trinity processors ?

[Hoonigan]

How about 2 or maybe 4 quad core Trinity processors ?

I can imagine that sort of setup would be extremely difficult to code for...

[CAT-THE-FIFTH]

It could be that the system has a quad core AMD APU with another GPU in Crossfire. The APU might be the successor to Trinity meaning it has GCN cores. I suspect that being a console the compute abilities of the GCN shaders might be put to good use. Look at Civilization V for example with both GCN and Fermi.

[Scribe]

I'm going to place my money on the Trinity A10-5800K, 3.8GHz with 4.2GHz turbo boost quad-core piledriver, Radeon HD 7660D with 384 cores at 800MHz, doesn't sound to bad to me Perhaps Sony could request a version that's further overclocked.

[CAT-THE-FIFTH]

I have a feeling GCN cores might be used. The compute abilities would fit the console market well IMHO. For example in some games which need less graphical effects the compute abilities of the shaders could be used to supplant the CPU.

[Noxvayl]

I would expect an x86 CPU to be a step back from the Cell processor the PS3 has, especially for gaming.

I'm not sure but I would suspect going from a Cell CPU to x86 would make backward compatibility difficult, something Sony have hopefully learnt is important to gamers.

[Percy1983]

Dare I say AMD did an amazing job with the Xenos by tweaking the architecture for a console, Sony probably figured they need the same AMD love.

Dare I say if you had the 360 GPU and PS3 CPU that would have been a more impressive machine.

[watercooled]

Any relatively modern x86 really wouldn't be a step back from the massively over-hyped Cell processor, it really isn't the supercomputer chip it's made out to be, it basically contains one of the cores the Xenon (Xbox CPU) has 3 of, with 7 (8 but 1 disabled for yield) simple MIPS cores which must receive instructions through the PPC core. And I doubt any x86 CPU would be even remotely hard to program for (and get a useful amount of performance from) compared to the Cell.

But I digress, I'd be very, very surprised if any of the new consoles used an x86 CPU; when carefully programmed for, a RISC CPU of some sort makes much more sense in a console. The only console I know of to have used an x86 CPU was the original Xbox which used nearly off-the-shelf components, but only because it was developed in a hurry. And you can't just drop a normal CPU into a console for other reasons, not least of which is DRM; the engineers want to make it as hard as possible, ideally impossible, for people to copy games/cheat/make unauthorised mods/load an unauthorised OS and so on. System buses can be easily sniffed so many are encrypted, even system memory is mostly encrypted. And then you have other features like efuses which prevent users installing custom firmware/accessing the JTAG interface. And so on...

I doubt it would be a high-end GPU of any sort, it's just not cost effective, would use too much power and would be difficult to cool quietly in the small cases consoles come in. I think the rumoured 6670 for the Xbox 720 sounds about right.

[Funkstar]

What kind of resolution do high end gamers on PCs use these days?

Consoles only need 1080p maximum. This used to be really difficult, now it's nothing special.

I can't see new consoles being the cutting edge in gaming like they once were, PCs can do that far better. But they will be entertainment systems that will hook into their manufacturers network of devices and (to steel an often used phrase) "just work". I'd be very surprised if a PS4 wasn't smaller, quieter and more efficient than the PS3. I'm not sure it would need to be too much more powerful. As long as it can do 1080p and 3D gaming that will do.

[watercooled]

On the current consoles, most games run at <720p (often far less) and are upscaled to the resolution they output by various methods, the 360 has a hardware scaler while the PS3 has only a horizontal scaler. A current mid-range GPU should allow reasonably-well-optimised games to run at native 1080p with reasonable settings with a playable framerate, quite a step up.

I'm really hoping that the next consoles have some thought put into idle power consumption, an area massively neglected by current consoles especially considering they're often used as media consumption devices and left on for extended periods, which would also help with heat + noise, also desirable for such a device.

[Scribe]

The CELL CPU was a great idea at release and if NVIDIA had done a better job on its end, the PS3's performance would have been epic.

Now though, that GPUs support general compute and are becoming more and more flexible in what tasks they can be given, there's a limited scenario for using the CELL SPEs as a lot of floating point number crunching can be handled by the GPU and for basic number crunching, GPUs do a better job because they have more, simpler cores, whilst the CELL sits somewhere in-between the two ends of the stick.

I strongly suspect that now the focus is to move towards making developer lives easier with unified memory and the ability to easily interchange between CPU and GPGPU cores, working on the same data. Not saying the PS4 will have this, but I suspect everyone will work towards it and I do believe there's a lot of untapped potential in learning to use GPGPU as second nature when writing a program.

[watercooled]

The CELL CPU was a great idea at release and if NVIDIA had done a better job on its end, the PS3's performance would have been epic.

Not really; the concept was a decent idea but the implementation was pretty badly flawed as I've briefly touched upon above. It's not fair to blame Nvidia for the amount of time+money wasted on the Cell, the GPU used was chosen by Sony, not Nvidia, and performance is similar to the Xenos.

I strongly suspect that now the focus is to move towards making developer lives easier with unified memory and the ability to easily interchange between CPU and GPGPU cores, working on the same data. Not saying the PS4 will have this, but I suspect everyone will work towards it and I do believe there's a lot of untapped potential in learning to use GPGPU as second nature when writing a program.

Most definitely, and a new console having a compute-enabled GPU would accelerate overall adoption of the technology, as has been shown in the past with SMP for instance; consoles, at least after a few years, generally start to struggle with unoptimised code, forcing the developers to use what's available to them.

[Scribe]

Not really; the concept was a decent idea but the implementation was pretty badly flawed as I've briefly touched upon above. It's not fair to blame Nvidia for the amount of time+money wasted on the Cell, the GPU used was chosen by Sony, not Nvidia, and performance is similar to the Xenos.

I would most certainly agree that x86 would have been easy to code for, there's a large experience base and only one type of core, having said that, for many games, the bottleneck is in floating-point calculations and being able to code the SPEs in low-level C as opposed to assembly, with programmable branch hints meant it was possible for developers to squeeze one heck of a lot of predictable performance out of SPEs without too much effort. The hard part is in balancing how to distribute the workload across them.