The New and Improved "G80 Rumours Thread" *DailyTech specs at #802*

And that raises interesting questions.
Maybe the drivers, when running under WinXP/DX9, can be tweaked to allow that limit to change, according to specific games.
Like SLI profiles, only for different, fixed ratios.

 

Now that idea is sexy!
If you could set the ratio yourself, than would be even sweeter.

 

Why (on a unified shading architecture) should DX9 apps use a fixed ratio between VS and PS?
I can't see any (obvious) good reason for that.

 

Maybe they don't want use resources on writing that part of the driver when DX9 games expect a fixed ratio between VS and PS anyway?

 

Or a fixed ratio gives better results than a load-balanced ratio...

 

There's dynamic load balancing on a USA--it doesn't matter what the game "expects," the USA will always give you the optimal result.

 

I think we should wait to find out what kind of load balancing NVidia has implemented...

Jawed

 

I seriously doubt that anyone would bother to implement an USA and not do any form of loadbalancing. But if it always gives the "optimal" result remains to be seen of course.

 

I enter this as evidence for my claim. DX9 doesn't care about the hardware, either--it'll work fine with a unified shader architecture, as evidenced by G965.

 

Load balancing and unified don't automatically go hand in hand - it certainly seems like the sensible thing to do, but it an implementation detail whether its there and and further to that, it can still be implemented in more than one fashion.

Unless, of course, you are publically confirming some things...

 

Of course he's not And it's not nice of you to ask, either

 

What about DX9 synthetic test. If the there is the ability to statically set the ratio would it not give results that have no relationship to "real word" performance? It would be pretty convenient for optimizations to be made if the workload is known.

 

Well, yeah, there was that Huddy comment that Xenos loses quite a bit of performance when it's using a fixed ALU ratio. so of course it's possible to have an architecture with a unified instruction set but a fixed ratio of ALUs.

 

(Stupid forum - too short a message my arse!)

 

The only obvious suggestion is that Load balancing is an element of the driver rather than on the chip. Which would imply some sort of course few times a frame load balancing if true.

But it's just as likely to be someone misunderstanding what they were told.

 

Why would nVidia use fixed ratio if it decreases performance? They don't have USA or they can't make it work properly?

 

Well, the oft-quoted and oft-misleading patent [backs up salt dumptruck] describes several load balancing schemes -- a fixed priority mode, a priority dependent on workload in vertex and pixel buffers, and a not-fully-detailed mode that appears (from what I could tell) to analyze the pixel-coverage of pending primitives.

If both G80 and Xenos are using priorities, rather than ALU allocation, I'm not sure how you can really get a fixed ALU ratio...? :shrug:

 

Sounds pretty amazing to me, but you make it sound like your disappointed.

 

My guess [for the fixed ratio] is that the automatic/dynamic load-balance [for G80 especially] is not feasible for the current generation of games, where vertex vs. pixel data ratio quite biased -- a single vertex could mean a hundred or a thousand pixels to be generated in the post-process stages following the pipeline. If the driver itself is responsible for handling the load-balance [thus, making it transparent for the rest of the system -- API and up] we still have to see, what will come out in the next few releases.
I bet for maximum 2 clusters dedicated for VS processing -- rest is for the fragments.

 

You're saying synthetic tests have any relationship to "real world" performance to begin with? A synthetic test will give you exactly what it claims: the maximum theoretical throughput of whatever subsystem assuming nothing else if happening in the system. So basically the result goes from meaningless to even more meaningless.