Embedded memory, daughter/parent die

ATI: So we have 48 shaders, each one of them does 4 floating-point ops per cycle, so 196 floating ops per clock.

On the embedded memory, I call it intelligent embedded memory or intelligent DRAM, because we actually have a lot of the [pauses] well the graphics pipeline continues into the memory. So we actually have a parent die and a daughter die on one substrate inside of the package. So if you look at the Xbox 360, and you open it up you’ll see that there’s one VPU package, but inside there are two different dies and the pipeline continues across those dies.

The part of the pipeline that’s in the daughter die, that lives with the memory is the z, the alpha, the stencil, the processors to do the resolve from multisample AA, the one that gives you AA for free, well 4x or 2x multisampling, and really what that does is it breaks one of the traditional bandwidth problems inside traditional graphics architectures. It keeps all that data so you don’t have to do a read or a write..Well you don’t have to read it into the general purpose, across the bus, process it and then write it back out to the memory. It all happens within the memory, and this just gives you phenomenally more bandwidth. And that’s how we get the anti-aliasing for free. The result for the anti-aliasing, you have 4 samples for every pixel, you ultimately end up with and the resolve is done right within that memory, so it doesn’t have to be read back into the parent die.

On the intelligent memory there’s something that we call internally the, well it’s a memory export path. So anywhere down the pipeline that you put data into the pipeline you can export that data out back to EDRAM or Level Two cache and then use that data again. So, and we really haven’t been able to do that before, so it allows you to do things like higher order surfaces for real this time, so it’s just another feature that gives us a lot of flexibility in the system.

FiringSquad: So the daughter die is where the embedded memory is?

ATI: Yes, the memory is there. There’s 10MB of embedded memory surrounded by “intelligence”. That’s the 192 component processors in there that do the work on the embedded memory data. There’s also a very high speed connection between the parent and daughter die, in case you do get have to get data back and forth, and that connection is, well there’s a 2GHz wide bus connection between them.