Is PS4's focus on Compute a proper substitute for a more powerful CPU?

[Arwin]

Geometry and materials, as was explicitly mentioned. Watched the first 34 mins and the MADD demo in the living room, glad we have a proper next gen experience at launch

 

[3dilettante]

That's exactly what I wanted to know. Will Mantle cover low level compute capability. I knew PS4 doesn't need Mantle per se. It has something "similar".

I'm not sure what choice was made, given the limited information disclosed.
It's not like AMD doesn't have a large compute initiative in parallel with its graphics pipeline, one which Sony has work going into.

I think it could be helpful for writing an engine that uses Direct Compute if Mantle rolls a path for lower-level compute in the same graphics framework for developers.
The other option might be run it with HSA providing basis for the compute portion.
Mantle does seem be borrowing some of the queuing methods for kernel dispatch used for AMD's lower-level compute methods, so they might be complementary.

The unknown here is that the latency numbers are given in the context of an HSA engine, which is the low-level access a Mantle-type compute path would try to get to.

 

[MBDF]

That's not compute replacing CPU. That's a programmable GPU doing graphics work.

Fair enough.

 

[patsu]

I'm not sure what choice was made, given the limited information disclosed.
It's not like AMD doesn't have a large compute initiative in parallel with its graphics pipeline, one which Sony has work going into.

I think it could be helpful for writing an engine that uses Direct Compute if Mantle rolls a path for lower-level compute in the same graphics framework for developers.
The other option might be run it with HSA providing basis for the compute portion.
Mantle does seem be borrowing some of the queuing methods for kernel dispatch used for AMD's lower-level compute methods, so they might be complementary.

The unknown here is that the latency numbers are given in the context of an HSA engine, which is the low-level access a Mantle-type compute path would try to get to.

I suspect sebbbi's posts are closer to the desired working model for the PS4's GPU. Essentially, the GPU becomes the "master" of the process. In this sense, the so called SPURS like scheduler would live in the GPU.

 

[Bigus Dickus]

Somewhat off topic, but looking at photomicrographs of ps4, xb1, kabini et al, and noting how tiny the jaguar cores really are, I wonder if there would be any benefit of AMD embedding say two jaguar cores into the gpu die (speaking of traditional discrete desktop parts) or having more closely integrated and designated cores in an APU type design? I know this seems counter to the gpgpu concept and perhaps seem redundant in an APU or SoC, but these cores would be there to work in tandem with CU on gpgpu tasks, perhaps allow devs to move certain game engine task programming into a more tightly controlled and predictable environment, etc.

I wouldn't have wondered if not for the truly small size if these cores.

 

[3dilettante]

Jaguar cores come in quads, and two of them are roughly 6mm2.
At 20nm, the difference would be small.
The desktop dual-core modules come out relatively close to the same size as the Jaguar ones, also.

There still are portions of frame processing and GPGPU operations that run better on narrower FPUs due to divergence or small granularity, and there are some that run equivalently (and would have the perverse outcome of providing a CPU to offload functions from the GPU), so there would be a computational benefit.
Perhaps putting them in poses other implementation concerns, or the benefits haven't outweighed the costs.

The rumored PCIe endpoint mode for Kaveri would have been just that.

 

[Bigus Dickus]

Hmm, interesting.

 

[Inuhanyou]

Saw this on another thread

http://gamingbolt.com/infamous-seco...ectly-on-ps4s-gpu-does-not-involve-cpu-at-all

“Physics simulation, collision detection, ray casting for audio, decompression and the like. And these operations are fine grained meaning that there will be many small world simulation tasks running on the GPU simultaneously alongside rendering of the game scenes. So the concept is that as game developers learn to use these techniques later on in the console life cycle, we will see richer and even more interactive worlds,” Mark Cerny said at the time. It seems that Sucker Punch’s inFamous Second will be using the above mentioned feature of the PlayStation 4. Brian Fleming, one of the founders of Sucker Punch studios revealed that the inFamous Second Son’s particle effects is being directly compiled on the graphics chip and uses no CPU processing at all.

So this would be like the PC physX implementation so far? I didn't even know Infamous was already using compute

 

[patsu]

It is "just" the particle system at the moment, but I would be thrilled if they use it for KZ4 large scale environmental effects. Skip the clean, clear blue sky levels.

 

[SlimJim]

Just thinking about it: to directly answer the thread title:
Yes: compute in a way is optimising. Optimising is always a proper substitute for a more powerful CPU.

 

[Arwin]

Wonder if we'll be able to see anything like this on consoles:

It actually seems as if Resogun's Voxel work could approximate this a bit.

 

[Deleted member 11852]

Wonder if we'll be able to see anything like this on consoles:

Someone somewhere is probably trying to register a trademark for Kinect Snowball Fight(tm).

 

[Shifty Geezer]

I imagine the computational power needed is mahoosive. I doubt a moderate sized GPU was driving that snow simulation, or anything close. More likely an insanely expensive compute farm was beavering away at the calculations, and quite possibly not in realtime (as CG work likes to pick visual targets that can't be rendered in realtime).

Yes, official pdf technical paper. The snowplough was the fastest anim to render at 2.1 seconds per frame. Although they aren't using GPGPU so they might find opportunity to accelerate that massively, we're not going to hit realtime, or anything like, for a good while yet.

 

[jlippo]

I imagine the computational power needed is mahoosive. I doubt a moderate sized GPU was driving that snow simulation, or anything close. More likely an insanely expensive compute farm was beavering away at the calculations, and quite possibly not in realtime (as CG work likes to pick visual targets that can't be rendered in realtime).

Yes, official pdf technical paper. The snowplough was the fastest anim to render at 2.1 seconds per frame. Although they aren't using GPGPU so they might find opportunity to accelerate that massively, we're not going to hit realtime, or anything like, for a good while yet.

That's 2.1 minutes/frame and only the simulation.

Pretty sure that voxel/particle hybrid simulations can be done properly with gpgpu and this new generation of consoles, but not as heavy version as we see in this paper.

 

[Arwin]

I'm sure we'll see.

 

[Shifty Geezer]

That's 2.1 minutes/frame and only the simulation.

Yeah, stoopid typo by me. Would need a 2.1x60x30(fps) = 3780x speedup. I don't think PS4's (or any) GPU can get a ~4000x speed increase over Disney Animation's render farm...

 

[Inuhanyou]

maybe in 10 years then...