PS3 GPU not fast enough.. yet?

[Magnum PI]

That's why he's the PI and we're not

well said !

remember kids.. being a PI is dangerous, it isn't a work meant for everyone.

and certainly none for the casual nerd whahaha !

this morning i saw the episode when i'm shot and i'm in a comatous state.. why friend rick in in jail, michelle has a new man.. so sad !

well.. gotta go as the men in white are coming for me.

 

[Titanio]

One thing I'll say is, it's quite funny seeing how it's being reported and interpreted by The Inquirer. I heard this stuff a while ago, and little 'hallelujahs' were being sung at the time with regards to this info in certain quarters! People seem to be focussing on the Cell bit, but the more important implication relates to RSX. It has pretty unfettered access to both pools of memory, Cell has unfettered access to XDR. What does that tell you? And why would they optimise it in that way? The suggestion is quite pleasing as far as RSX is concerned. If they were to optimise FlexIO in one way, RSX was certainly the right client to prioritise.

The only implication for Cell is that it has limited access to GDDR3 (and not also XDR as some have claimed - Cell has full and proper access to XDR). If data is to be shared back to Cell from RSX or GDDR3, RSX should do the pushing (with its 15GB/s) rather than Cell doing the pulling (16MB/s).

 

[Magnum PI]

EXIF data is actually quite easy to change.

that's precisely because of that that i say "according to EXIF data".

 

[aldo]

according to EXIF data the picture was taken the 4th of June, 20:20:59

Charlie Demerjian at inq, who wrote the article, emailed me a little over 3 weeks ago about this coming bombshell(?). He seemed extremely excited about this scoop and said he was going to release the story the week of the 14th. So the picture source is no doubt 3+ weeks old.

inq - the source was saying... that the RSX can only write about half as much vertex data as it can fetch, not an ideal situation by any stretch, but survivable. Then came the horrible news, RSX appears to be limited to setting up 275 Million triangles/second

I know it's inq, but anyone concerned about this statement?

-aldo

 

[aldo]

LOL!

Joystiq responds to The Inquirer's article:

The Inquirer, is running a piece titled, "PS3 hardware slow and broken." It's full of the usual Inquirer stuff: bad spelling, horrible grammar, and (most importantly) specious arguments. So, before we even get to what they're claiming, understand that anything published on The Inquirer should be considered entertainment, much the way the Enquirer is.

-aldo

Edit: Fixed Joystiq link. Thanks for the heads-up and fix Dave.

 

[Carl B]

Well, Joystiq's refutation doesn't really amount to much since it links to Cell-centric, rather than PS3-centric, articles as reference...

But they're of course right in denouncing Charlie D's own logic as totally off the wall.

 

[Titanio]

The base data is correct, it's the interpretation layered on top of that and conclusions reached that are FUBARed. Demerjian was seeing 'bombshells' where there are none.

As for the triangle setup rate, it's much higher than you're ever going to be typically processing. I don't think I've ever heard of someone being bound by their tri setup rate. I don't want to write it off totally, someone might chirp in with situations where a high setup rate may be desireable. But given a regular per-vertex workload and a regular amount of hidden triangle removal (pre-GPU), you'll probably bottleneck on something else sooner than the setup rate if you try to actually draw that number of triangles!

edit - to give you an idea, I was just looking at one ATi presentation about general GPU optimisation, and under Triangle Setup it simply lists that it is never a bottleneck.

 

[Dave Glue]

Fixed joystiq link:

http://www.joystiq.com/2006/06/05/rumor-ps3-hardware-slow-and-broken/

 

[ERP]

The base data is correct, it's the interpretation layered on top of that and conclusions reached that are FUBARed. Demerjian was seeing 'bombshells' where there are none.

As for the triangle setup rate, it's much higher than you're ever going to be typically processing. I don't think I've ever heard of someone being bound by their tri setup rate. I don't want to write it off totally, someone might chirp in with situations where a high setup rate may be desireable. But given a regular per-vertex workload and a regular amount of hidden triangle removal (pre-GPU), you'll probably bottleneck on something else sooner than the setup rate if you try to actually draw that number of triangles!

edit - to give you an idea, I was just looking at one ATi presentation about general GPU optimisation, and under Triangle Setup it simply lists that it is never a bottleneck.

I'd be willing to bet you'll never get close to the 275 million number on a PS3 in any real application.

 

[ROG27]

I'd be willing to bet you'll never get close to the 275 million number on a PS3 in any real application.

I'll second that. Inquirer's umpteenth goof-up. ******ism exposed (SHOCK!). It is extremely sad seeing this entity trying to pass itself off as an unbiased and legitimate hypermedia publication.

 

[Acert93]

As for the triangle setup rate, it's much higher than you're ever going to be typically processing. I don't think I've ever heard of someone being bound by their tri setup rate. I don't want to write it off totally, someone might chirp in with situations where a high setup rate may be desireable. But given a regular per-vertex workload and a regular amount of hidden triangle removal (pre-GPU), you'll probably bottleneck on something else sooner than the setup rate if you try to actually draw that number of triangles!

I actually remember a few developers here making negative comments in regards to the triangle abilities not being as high, relatively, this gen compared to last. This and the issue of "small triangles" have actually been mentioned quite frequently on the forums.

I would not mind some developers giving some more input on these areas. Since vertex processing tends to be bursty, how does the peak setup limit play into this? If you spend 15% of your time doing most of your vertex work will this be an issue or more of a meaningly side note?

I don't mind bump mapping techniques, but I am hoping Cell on the PS3 and datastreaming/tesselation/unified shaders on the 360 allow more flexibility in game design so we see more games with larger poly budgets.

Actually... a new thread on, "Polygons on Next Gen Consoles" would be a good/interesting thread if we could get some developer feedback. Any indepth discussion here is bound to scare away meaningful comments due to the title.

 

[Carl B]

I'll second that. Inquirer's umpteenth goof-up. ******ism exposed (SHOCK!). It is extremely sad seeing this entity trying to pass itself off as an unbiased and legitimate hypermedia publication.

Charlie D in particular... I mean I'm not one for conspiracy theories, but this guy just has a naked agenda. Article after article from him...

It's one thing to have an agenda, but couple it to his lack of understanding and the reach of The Inquirer - and it's just a formula for the spread of ignorance.

 

[Titanio]

I actually remember a few developers here making negative comments in regards to the triangle abilities not being as high, relatively, this gen compared to last. This and the issue of "small triangles" have actually been mentioned quite frequently on the forums.

I would not mind some developers giving some more input on these areas. Since vertex processing tends to be bursty, how does the peak setup limit play into this? If you spend 15% of your time doing most of your vertex work will this be an issue or more of a meaningly side note?

I don't mind bump mapping techniques, but I am hoping Cell on the PS3 and datastreaming/tesselation/unified shaders on the 360 allow more flexibility in game design so we see more games with larger poly budgets.

Actually... a new thread on, "Polygons on Next Gen Consoles" would be a good/interesting thread if we could get some developer feedback. Any indepth discussion here is bound to scare away meaningful comments due to the title.

I'd question how much resources you'd have to spend per vertex or whatever if you actually tried to draw such a number of triangles on the screen. The only comment I can remember from a dev in this regard was actually ERP's original comment about the level of shading you'd be reduced to if you tried to max Xenos's. I think other things would become a problem much sooner than your setup rate if you actually tried to draw so many such that the setup rate itself would become an issue. I'd also wonder if it'd be possible at all if you have in any way decent HSR going on on the CPU..I mean if you had 4.5m visible triangles per frame or whatever, you'd be talking about sub-pixel triangles!

That thread might not be a bad idea though! The issue of bursty vertex work and how that would relate to setup rate, or indeed how one could smooth vertex work out over a frame for better utilisation in a fixed ratio architecture could be interesting points of discussion.

 

[Dave Baumann]

Bear in mind that some passes are very geometry heavy passes (i.e. a Z setup or shadow passes).

 

[Tap In]

so wait let me get this straight... the true facts about this info are:

1) RSX triangle set-up is 275,000
2) Cell can not reasonably read from GDDR
3) The Inq are a bunch of over-the-top crazy numbskulls

 

[Titanio]

so wait let me get this straight... the true facts about this info are:

1) RSX triangle set-up is 275,000

275m according to the report..it's not clear if that is at the current clocks in the dev kits or final clocks.

2) Cell can not reasonably read from GDDR

Not directly. If you want Cell to access something from GDDR3, you get RSX to write that data to Cell or XDR - RSX reigns over FlexIO basically, and that is a good thing.

3) The Inq are a bunch of over-the-top crazy numbskulls

True The numbers are all right, but their interpretation is over-the-top crazy.

 

[ShootMyMonkey]

I mean if you had 4.5m visible triangles per frame or whatever, you'd be talking about sub-pixel triangles!

Or lots of overdraw. Or lots of render passes involving the same (source) geometry over and over (e.g. shadow passes).

 

[Mmmkay]

What about the 'RSX acts as a southbridge' stuff? That's quite a departure from what we had been originally hearing isn't it?

 

[Tap In]

275m according to the report..it's not clear if that is at the current clocks in the dev kits or final clocks.



Not directly. If you want Cell to access something from GDDR3, you get RSX to write that data to Cell or XDR - RSX reigns over FlexIO basically, and that is a good thing.



True The numbers are all right, but their interpretation is over-the-top crazy.

thanks T

it's a shame someone couldn't take this info and write something intelligent about the design and how it actully works in real world, instead of turning it into a ridiculous headline.

Now I'm mad I even gave them a hit to their web page.

 

[Dave Baumann]

275m according to the report..it's not clear if that is at the current clocks in the dev kits or final clocks.

It wouldn't make sense if it wasn't for the final clocks. Its two clocks per triangle.