It handles all the control flow.
For all we know...is Rys correct?.
- Thread starter Love_In_Rio
- Start date
It handles all the control flow.
It is used in texture processing.
They read the resource constants and cache them, allowing the GPU to support any number textures per shader. Same as that "Bindless Texture" that was anounced with Keplar, we just didn't reference it because we have PRT.
That's understandable as there is none.
"SIMT" (a really ridiculous term in my opinion) is just a stupid way to describe the mapping of a problem formulated in an implicitly parallel way onto an SIMD architecture with lane masking support.
You have to admit though, that SIMT is a bit more catchy than "tmoapfiaipwoaSawlms" [my bold above].
In retrospect, this wasn't a great choice IMHO, as one is not necessarily and immediately obviously a superset of the other, and the bindless thing is pretty damn useful.
We can't really make a full GCN v Kepler comparison until nVidia rolls out its Tesla parts. GK104 is 40% bigger than Pitcairn and 40% faster so that's a wash. 40% with only a 25% bandwidth advantage too.
Not really sure what defines architectural "elegance" anyway. Doing more with less? Prettier slides?
]Dave Baumann said:
I'm afraid I lost you at the PRT part.
I think Dave means that you can achieve a pretty similar effect on hw having PRT...think instead of having >128 bind points, thus having to rely on the bindless aspect, you just make a fat, partially resident thingie and index into that (at least that's something that took about 30 seconds to cook up - probably wrong too!). You get the benefits of the CPU savings (no messing up with setting and unsetting resources, which isn't cheap), and some hardware help. As I said, if that's the idea, it's not a complete superset or a seamless replacement for the bindless facilities, so the fact that they can do it yet chose against adevertising it was counterproductive from where I'm standing.
I take it Partially Resident Textures make bindless ones possible too.
Sure, mention it to the people that are going to get the most benefit (i.e developers), but at "Tech Day" you are apealing to the end users as much as press. If I have a few features that are interesting but outside of the DX model at the moment I'll spend time talking about the one that is more likely to resonate with the users ("Oooo, hardware accellerated MegaTextures").
Not only, it also can improve performance in the CPU.Davros said:
In the 7970 block diagram, the Ultra-Thread Dispatch Processor, which had been around from the X1800 to the 6970, is missing. I know each GCN unit has a scheduler, but that's at a different level I think.
So what happened to the thread dispatcher? Is it still there but they just didn't show in the block diagram, or they don't need it anymore? If so, how is work being dispatched to all the GCN units? Is it the ACE units?
So what happened to the thread dispatcher? Is it still there but they just didn't show in the block diagram, or they don't need it anymore? If so, how is work being dispatched to all the GCN units? Is it the ACE units?
That's over-specialisation IMO. What you probably want is "Oooo, shiny!!", and saying that you support infinite anything, in this case textures, gets it. I doubt that the vast majority of users tick upon reading MegaTexture. Maybe it's just me. Also, NV seems to have done quite well in having users drool over bindless. It gave them fuel to present something hyper novel, instead of something that ATI also does. That's not to be scoffed at, is it?
Similar threads
