SIGGRAPH paper on Larrabee architecture

[sonoran]

Some Larrabee info has just been made public. A few links for anyone interested...

Quick Intel Fact Sheet on Larrabee: http://www.intel.com/pressroom...eleasepri_20080804fact

A link to the SIGGRAPH article on the Larrabee architecture: http://doi.acm.org/10.1145/1360612.1360617

 

[CTho9305]

Interesting read. I wonder why only Kyro (and Intel, who I think licenses their stuff) does tile-based rendering. There must be another side to the story...

 

[TuxDave]

Originally posted by: CTho9305
Interesting read. I wonder why only Kyro (and Intel, who I think licenses their stuff) does tile-based rendering. There must be another side to the story...

It says in the paper that the tiles are designed to fit within a single core's cache so that it obtains all the information it needs through one memory operation. What arguments do you see against it?

 

[CTho9305]

Originally posted by: TuxDave

Originally posted by: CTho9305
Interesting read. I wonder why only Kyro (and Intel, who I think licenses their stuff) does tile-based rendering. There must be another side to the story...

It says in the paper that the tiles are designed to fit within a single core's cache so that it obtains all the information it needs through one memory operation. What arguments do you see against it?

I assume nvidia and ati aren't both stupid...

 

[TuxDave]

Originally posted by: CTho9305

Originally posted by: TuxDave

Originally posted by: CTho9305
Interesting read. I wonder why only Kyro (and Intel, who I think licenses their stuff) does tile-based rendering. There must be another side to the story...

It says in the paper that the tiles are designed to fit within a single core's cache so that it obtains all the information it needs through one memory operation. What arguments do you see against it?

I assume nvidia and ati aren't both stupid...

I was hoping for a pro vs con argument for tile based rendering.

(note this is just my guess and not based on any facts 😛 )
Something like: tile-based and non-tile based rendering can both do graphics rendering with maybe different performance levels. However the same hardware that makes tile based rendering good is also good for highly parallel scientific computing which is why they chose one style over the other. A graphics company who's #1 priority is graphics rendering may have found a different architecture faster for what their core business is but may lose the flexibility of other computational workloads.