Xenos as Physics Processor?

[ihamoitc2005]

Maybe

I think the point is those shaders aren't outputting per instruction, but applying multiple instructions to achieve better shading. On simple shaders the ROPs wouldn't be enough, but on complex shaders they're plenty. And anyone using only simple shaders on next-gen games is going to end up with inferior looking titles

On the flip side if you had enough ROP's to satisfy the simplest, faster shaders outputting at full speed, most of the time they'd be sitting idle.

You are right that if 48 shaders are double-strength then you have faster performance from complex shaders as you would with any doubled up shader unit but with only 8 ROPS even not so complex shaders can = bottle-neck and simple shaders will = extreme-bottleneck therefore it is not logical for 48 double strength shaders no? Doubled up US = major bottle-neck and inefficiency in all but most complex shader situations. Given chip size, supporting components, need for MSAA components, it is only logical that Xenos = 48 single-strength shaders. Too much supporting evidence for this view, including Microsoft article comparing PS3 to Xbox360, and no evidence to support double-strength unified shader theory.

G70
24ps:16rops = 1.5:1 ratio
Xenos possibility a (single strength)
48/2 US:8rops = 3:1 ratio
Xenos possibility b (double strength)
48 US:8rops = 6:1 ratio

 

[Shifty Geezer]

Doesn't texturing also tie up a G70 shader though? So in conventional use those shader's are going to be used very fully, and when not texturing to turn those resources to pixel shading. Assuming I understand the architecture right and it's likely I dont.

 

[Jawed]

http://www.hardspell.com/newsimage/2005-6-21-16-10-14-654986702.gif

Do the maths from the above table and compare to diagram below,

http://www.hardocp.com/images/articles/1119063771Y3O0GyEDBw_3_3_l.jpg

The scalar flops are from the mini-ALUs...

Yes, you're right, the G70 and RSX GFLOPs counts look like they include the mini-ALUs as scalars.

So 313GFLOPs in 7800GTX versus 170GFLOPs in X1800XT amounts to roughly equal performance? Agreed?

I hope that's the last we hear of GFLOPs from you Jaws, if that doesn't put the final nail in the coffin then nothing will.

Xenos may well not have a mini-ALU with the same functionality as seen in R300...R420...R520. But I can't see why not. If it's valid in those architectures, why isn't it valid in Xenos?

Overall, in my view it's safer to ignore the mini-ALUs entirely, because you don't know what the instructions and dependency limitations are.

They wouldn't be called mini-ALUs if they were full function. If the MUL capability of a mini-ALU is limited to multiplying/dividing by 2, 4 or 8 is that a general purpose MUL?

Jawed

 

[Acert93]

7800GTX is 313GFLOPs? It is more powerful than CELL!

New theory: PS3 will be 2x 7800GTX @ 550MHz. You heard it here folks, the PS3 is really an SLI rig hehehe

Oh, architecture you say... (Ok, bad attempt at humor... back to the shark fight... or is it a shark vs. a short man with a killer hat?)

Anyoooooow... how 'bout the GPGPU / Physics stuff found in this thread?

Xenos as Physics Processor?

Ps- a thread dedicated to GPU architecture would probably be good. Separating the Physics stuff and the general performance/effeciency questions would probably be neater and allow more focused discussion. I understand how they dovetail and the R520 brings out some new questions, but the thrust of the two discussions is a little different.

 

[Shifty Geezer]

Maybe Jaw's'/'ed' could have their own thread too, to keep the same arguments from pervading any debate with Xenos in it?

 

[Nemo80]

Anyoooooow... how 'bout the GPGPU / Physics stuff found in this thread?

Xenos as Physics Processor?

Ps- a thread dedicated to GPU architecture would probably be good. Separating the Physics stuff and the general performance/effeciency questions would probably be neater and allow more focused discussion. I understand how they dovetail and the R520 brings out some new questions, but the thrust of the two discussions is a little different.

Utter BS.

 

[Mefisutoferesu]

Acert, you're thinking of the tall guy with metal teeth. The shortguy with the killer hat is Oddjob. Wait, is he even in this forum? Maybe it was GAF? TeamXbox? Well, anyway, one way or another the 1st person to use "JamesBond007" in this forum is royaly screwed.

 

[Acert93]

Acert, you're thinking of the tall guy with metal teeth. The shortguy with the killer hat is Oddjob. Wait, is he even in this forum? Maybe it was GAF? TeamXbox? Well, anyway, one way or another the 1st person to use "JamesBond007" in this forum is royaly screwed.

ROTFLOL... oh I cannot breath... help.... help.... <gasp!>

Thanks Mefi Now that everyone had a good laugh we can get back to talking about Xenos and Physics and getting a sticky Xenos/RSX thread up where the fishies can swim.

Btw, this same issue is being discussed in the PC GPU realm here . Appears MS has started work on a Physics API, specifically with comments about accelerating physics on GPUs.

 

[Nemo80]

Appears MS has started work on a Physics API, specifically with comments about accelerating physics on GPUs.

No it has not. The AGEIA library e.g. is already finished, and runs entirely on the 360 CPU because it's much better capable of handling this task.

 

[Acert93]

No it has not.

According to DeanoC they had already put out requests for programmers for a DirectPhysics API and the job position description specifically mentions accellerating physics on GPUs. If you had checked the link I created in that thread it is all spelled out.

So yes, MS has started work on a DirectPhysics API.

Your comments about Ageia are irrelevant and have no connection with what I have said. If you are going to disagree and toss around pointless comments like "BS" you better has something relevant to say.

 

[Dr. Nick]

We have discussed this a little, off and on. As for real world implications I think some time is needed. But the recent ATI Toy Movie maybe (?) a small implication of where some of this is going... exciting times.

If the Xbox360 can churn out graphics like the ones in that demo it doesn't have anything to worry about. After viewing the video on an HDTV even with the artifacting it doesn't look like anything thing this gen systems can do. I know it has been said many times in the past to describe other tech demos but this one looks like a somewhat low rent CG. Not exactly lord of the rings but it looks amazing none the less.

 

[PurplePigeon]

No it has not. The AGEIA library e.g. is already finished, and runs entirely on the 360 CPU because it's much better capable of handling this task.

If you log on to Microsoft careers, you can easily find job listings to work on DirectPhysics for GPUs and other hardware:

The Windows Graphics and Gaming Technology group is looking for a software design engineer to join a growing team responsible for developing Direct Physics. This team is responsible for delivering a great leap forwards in the way game developers think about integrating Physics into their engines. Physics and real time, accurate simulation is a key part of the next generation gaming experience, bringing increased realism, greater immersion and more interesting experiences.
As a member of the core engine team you will be responsible for working on technologies such as the core simulation engine, collision detection, animation, particle effects, dynamics and constraint solving. You will be primarily responsible for key components in the architecture, developing and optimizing the performance and memory footprint while helping scale the architecture across multiple cores and well as mapping the algorithms onto GPU’s and other high performance hardware. Extensive knowledge of performance optimization on the Windows platform is required to be successful and a good understanding of the needs and requirements of game developers would be useful.
You should have excellent C/C++ skills with at least 4 years of development experience. Knowledge of DirectX and shipping game applications is strongly recommended. An understanding of Havok, Ageia, MathEngine, Meqon or ODE would be ideal. A BA/BS degree in Computer Science, Math, Physics or equivalent experience is required.

 

[Nemo80]

If you log on to Microsoft careers, you can easily find job listings to work on DirectPhysics for GPUs and other hardware:

I'm talking about the X360, not some Direct X 10.

 

[Carl B]

No no, but this isn't DirectX 10 - this is complementary to it. And it just goes with Microsoft's philosophy in so many areas; offer a free product that's 'good enough' and people will flock to it. In this case they can start to build their dominion of the physics API space and thus make development on Vista, 360 - whatever - all the more inexpensive/appealing to developers.

And it would be foolish to think that at least some of what we see with the advent of DX10 won't trickle down to the 360. I'm sure the same can be said for this physics API.

It puts Sony - already made overnight with PS3 the almost de facto standard bearer for OpenGL - in the position where they may need to look to acquire a physics API or help to establish an open-source foil to Microsoft's push.

 

[Acert93]

I'm talking about the X360, not some Direct X 10.

Your response was to what I said that specifically mentioned MS developing an API in the PC GPU space:

Btw, this same issue is being discussed in the PC GPU realm here . Appears MS has started work on a Physics API, specifically with comments about accelerating physics on GPUs

Your comments telling me "you are wrong" take what I said out of context and used nonsensical point about Ageia and the 360 as proof (and for the record, Ageia is not the only Physics API out there; it actually is not even the most popular). Proof of what I am not sure since I never mentioned Ageia or the 360 specifically

And as XD said, it would be a case of tunnel vision to ignore that MS is intentionally leveraging the 360<>PC, and it is not a coincidence that one of their advertised features in the 360, MEMEXPORT, is designed to handle the VERY task their new Physics API is partly aimed at.

It is not a coincidence that Xenos has MEMEXPORT and that MS is developing a PhysicsAPI that can be accelerated on a number of devices including, but not limited to, CPUs, PPUs, and even GPUs. MEMEXPORT is not a standard feature, it is something MS asked for. Something to keep in mind.

Finally, comparing the MS API to Novodex Suit or the Havok Suit may be VERY premature. The MS API may be a thin software layer that is meant to direct the Physics Plugins to the desired target (CPU, GPU, PPU).

So it may be nothing more than an application that allows acceleration of Physics on PCs through any number of devices. Or it could be a full application. MS goes gone different routes in DX (Direct 3D, Direct Draw, Direct Sound, Direct Play, etc). It is not too hard to imagine DirectPhysics as DX and Novodex/Havok being the equivalent of UE3, D3 Engine, Source, etc. But we don't know yet.

Anyhow, if you were talking about the console only you misquoted me and are making a disagreement out of that, which makes no sense to me.

 

[Carl B]

Finally, comparing the MS API to Novodex Suit or the Havok Suit may be VERY premature. The MS API may be a thin software layer that is meant to direct the Physics Plugins to the desired target (CPU, GPU, PPU).

So it may be nothing more than an application that allows acceleration of Physics on PCs through any number of devices. Or it could be a full application. MS goes gone different routes in DX (Direct 3D, Direct Draw, Direct Sound, Direct Play, etc). It is not too hard to imagine DirectPhysics as DX and Novodex/Havok being the equivalent of UE3, D3 Engine, Source, etc. But we don't know yet.

Good points Acert, I may have been thinking too grandiose with my own predictions; indeed they might just tie it in as a means of allowing the existing engine makers to more easily access latent resources in a Vista environment. Ah well we'll see. Either way, another seemingly smart play by Microsoft in the development space.

 

[j^aws]

Yes, you're right, the G70 and RSX GFLOPs counts look like they include the mini-ALUs as scalars.

Well thank you.

So 313GFLOPs in 7800GTX versus 170GFLOPs in X1800XT amounts to roughly equal performance? Agreed?

Sorry but you're still stating mixed 16bit+32bit flops after it's been mentioned several times in this thread. Just counting 32bits, i.e. excluding 16bit normalise, it's ~ 230 Gflops peak for 7800GTX and it would be in the same ballpark as the X1800XT.

I hope that's the last we hear of GFLOPs from you Jaws, if that doesn't put the final nail in the coffin then nothing will.

You'll hear whatever is required from me to make my point. If you have trouble understanding my point, it has nothing to do with GFlops. If I need to use a particular metric, whether it's instructions, components, shaders, dots, flops, bananas, whatever, I'll use it.

Xenos may well not have a mini-ALU with the same functionality as seen in R300...R420...R520. But I can't see why not. If it's valid in those architectures, why isn't it valid in Xenos?

Because Xenos uses simpler ALUs but lots of them. More RISC like than CISC like. I'd imagine it would be even harder to schedule instructions with mini ALUs in a 16-way SIMD engine.

Overall, in my view it's safer to ignore the mini-ALUs entirely, because you don't know what the instructions and dependency limitations are.

They wouldn't be called mini-ALUs if they were full function. If the MUL capability of a mini-ALU is limited to multiplying/dividing by 2, 4 or 8 is that a general purpose MUL?

Off-course it depends on context...but a basic component ops/sec would give a good ballpark, assuming same precision 'bits'...