Performance: Pixel Pipelines and Speed

[Acert93]

I thought <A HREF="http://www.anandtech.com/video/showdoc.aspx?i=2290&p=8" target="_new">this</A> was very interesting. Since GPUs have developed into very powerful units utilizing a parallel structure to produce results, I found it really exciting that losing 30% in speed could be made up by adding 30% more pixel pipelines. This is REALLY effecient. The exciting part is this means as die sizes shrink every 18-36 months we will continue to see significant speed bumps with GPUs. This is good news Now if they could only find a magic solution for memory bandwidth...

Here is an interesting question: What can we expect from the R5xx and NV5x series? Will we see 24 or 32 pipeline parts? Or will we see more significant speed bumps? Any thoughts? My guess is ATi may go with a unified shader setup... at least that is what the leaked Xbox2 specs looked like.

 

[sweatlaserxp]

losing 30% in speed could be made up by adding 30% more pixel pipelines

So, are you saying that the 9700 Pro is twice as powerful as the 5800 Ultra?

<A HREF="http://atomfilms.shockwave.com/landing/landingIndex.jsp?id=dumb01&mature=accept" target="_new">DumbLand</A>

 

[Acert93]

Uhhh, no. Please read the link first. I was talking about a tradeoff within the same design. i.e. A 12 pipe and 16 pipe X800. Anyhow, I thought it was interesting--I did not expect this much effeciency. I think it boads well for the future...

 

[eden]

Consider this, your average CPU has not more than 9 pipelines, or even 6 in the P4's case, and without cache has about 40 million transistors. Amazing when you think of the amount of pipelines for the GFX cards. They must cost a lot of space! So 24 or 32 will make the die seriously huge.

--
<font color=blue>Ede</font color=blue>