RSX is a faster 7800GTX

[London Geezer]

I thought by competing, he meant that there would be a competition between pixel and vertex operations. He mentions load balancing in the following sentence.

Vertex and pixel operations already are competing in current GPUs.

Either the pixel shaders are waiting because the vertex shaders can't keep up - or the vertex shaders can't push any more triangles into the rasteriser because the pixel shaders can't keep up.

Xenos provides an automatic load balancer between the two - so that no ALUs go unused - they're always fully assigned to the current workload, whatever the relative proportions of the vertex and pixel workloads.

The developers can fine-tune the workloads by providing compiler hints about the typical amount of work in, for example, a loop. If the scheduler knows that a loop typically runs up to four times, it can use that information to bias scheduling towards looping 1 to 4 times...

Jawed

Yes that's the theory.

 

[Jawed]

Well luckily for ATI, current GPUs are performing so badly in this respect it'll be easy for Xenos to rung rings around them.

Jawed

 

[BenQ]

Well luckily for ATI, current GPUs are performing so badly in this respect it'll be easy for Xenos to rung rings around them.

Jawed

I read about how Unified shaders will be 95-100% efficient as far as keeping all the ALU's working all the time, but how inefficient are standard shaders?

If Unified shaders are 95-100% efficient then what percentage would standard shaders be sitting at?

 

[Shifty Geezer]

Well luckily for ATI, current GPUs are performing so badly in this respect it'll be easy for Xenos to rung rings around them.

Have you any metrics for current use?

 

[Jawed]

Do some searching - there's a vast amount written on this subject already.

Jawed

 

[Shifty Geezer]

I'm not that interested to go looking! But if you had a quick link I'd like a nose.

 

[Jawed]

Well luckily for ATI, current GPUs are performing so badly in this respect it'll be easy for Xenos to rung rings around them.

Have you any metrics for current use?

Only ATI's metric of 50-70% efficiency. They're prolly referring to the utilisation of a pipe, rather than the utilisation of ALUs - but it's vague.

Jawed

 

[Jawed]

I'm not that interested to go looking! But if you had a quick link I'd like a nose.

Sorry, if you can't be bothered, why should I?

But anyway, I was answering BenQ.

Jawed

 

[Shifty Geezer]

You made the assertion that current utilisation was poor. I presumed you had seen evidence for this and might know where to find it.

 

[3roxor]

We'll see soon enough which gpu has better real life performance.

 

[Jawed]

Well here's a simple example: in D3 and other games, one or more rendering passes (one per light) are made just to generate shadowing effects.

While those rendering passes are running, the pixel shader pipelines are doing nothing - there are no pixels to shade - the game engine is simply trying to work out the limits of each object and how the shadows fall.

So,. during each pass in RSX you've got 8 vertex pipes running, but 24 pixel pipes sat around twiddling their thumbs.

At least the 16 ROPs are doing something

Jawed

 

[Hardknock]

Xbox 360 GPU

Just to make things a bit more interesting, check out this video for "Full Auto" (Xbox 360):

http://www.xboxyde.com/leech_1486_1_en.html

In the video the developer confirms the game will have HDR Lighting + Full Scene AA. I remember a lot of people were wondering if it was possible to have both on the Xbox 360's GPU due to bandwidth issues. Looks like it won't be a problem, so it appears ATi has the one up on Nvidia in this area.

 

[AlNom]

Vertex and pixel operations already are competing in current GPUs.

No need to get all... like that. I'm just saying that he's talking about competition in the load balancing sense. Not company competition or whatever.

With unified=shared=competing pipelines, I thought he meant that problems would be from trying to do pixel shading and vertex shading at the same time. (Although I'm assuming you could do that in the first place )

 

[Fafalada]

So,. during each pass in RSX you've got 8 vertex pipes running, but 24 pixel pipes sat around twiddling their thumbs.

I like Unified shaders, but yours isn't an example where they provide any particular advantage.
The volume shadow scenario is 99% pixel output limited (bandwith or ROPs depending on your GPU), so shaders will mostly be sitting idle twidling their thumbs regardless.
We're also talking pretty simple vertex processing, so that 500Mpoly/s vertex setup will get saturated by only a few VSs as well (for the cases where pixel processing isn't the limit).

 

[Jawed]

D3's pointy head syndrome indicates that D3 is poly limited because the shadowing algorithm is expensive.

If the shadowing algorithm wasn't so expensive then there'd be less pointy heads and there'd also be more monsters on screen at the same time.

Jawed

 

[Fafalada]

D3's pointy head syndrome indicates that D3 is poly limited because the shadowing algorithm is expensive.

Volume shadowing is expensive because it renders very large polygons - requiring a lot of fillrate. If you prefer, the shadow polygons take a long time to render.
The vertex processing for the shadowvolume polygons - Including silhoutte tracing and volume extrusion - is trivial compared to the pixel fill costs.
I'm not making assumptions about how volumes perform here, this is based on hard numbers.

 

[Jawed]

The number of pixels for a square head and a square head with rounded-off corners is approximately the same.

A doubling in the poly count for the head wouldn't create a doubling in fill-rate required.

Jawed

 

[London Geezer]

The number of pixels for a square head and a square head with rounded-off corners is approximately the same.

A doubling in the poly count for the head wouldn't create a doubling in fill-rate required.

Jawed

There is a very well known mod that makes the heads perfectly rounded, with no loss in performance. That always made me wonder why exactly DOOM3 ever had pointy heads...

 

[one]

As for devs writing vertex work on CPUs...well, large sack of salt of course.

It reminded me of this Xenon patent.
http://www.beyond3d.com/forum/viewtopic.php?p=440265

 

[seismologist]

So,. during each pass in RSX you've got 8 vertex pipes running, but 24 pixel pipes sat around twiddling their thumbs.

I like Unified shaders, but yours isn't an example where they provide any particular advantage.
The volume shadow scenario is 99% pixel output limited (bandwith or ROPs depending on your GPU), so shaders will mostly be sitting idle twidling their thumbs regardless.
We're also talking pretty simple vertex processing, so that 500Mpoly/s vertex setup will get saturated by only a few VSs as well (for the cases where pixel processing isn't the limit).

That's some good info. If that's the case, then you wouldn't want your pixel limited output interrupted by spurious vertext shading requests. Is that what happens with the unified shaders?

Then it makes sense that some devs want to move the vertex operations onto the CPU then.