Next Gen Intel Gfx ( PowerVR 4 ?? ) - SM3.0 and HW T&L

[Ailuros]

it's a deferred renderer. and yes, it's ps2.0 and cpu emulated vertex shaders, but there's nothing wrong with that.

ed: just to clarify - by current i mean the gma 900/950.

If you want to call it a DR (on whatever weird basis) be my guest. I'll just call it then a DR that sucks

 

[darkblu]

If you want to call it a DR (on whatever weird basis) be my guest. I'll just call it then a DR that sucks

i call it a deferred renderer on the basis that it's a deferred renderer. and unless you know something that intel are hiding from the developers, i see no reason why you would call it differently. BTW, just to humor you, the fillrate of their IGP is a tad beyond what the bandwidth of their north bridge can deliver (given their IGP was an IMR). so intel must have produced one really dumb IMR part this time, so dumb that it cannot actually deliver on its fillrate.

 

[Ailuros]

i call it a deferred renderer on the basis that it's a deferred renderer. and unless you know something that intel are hiding from the developers, i see no reason why you would call it differently. BTW, just to humor you, the fillrate of their IGP is a tad beyond what the bandwidth of their north bridge can deliver (given their IGP was an IMR). so intel must have produced one really dumb IMR part this time, so dumb that it cannot actually deliver on its fillrate.

Intel integrated graphics solutions have had tile-based rendering for ages, and these are not based on PowerVR technology. Unlike the PowerVR solutions, they don't appear to have any kind of efficient overdraw removal, so performance and efficiency are generally not very good. IIRC, you can even run the Intel solutions in immediate mode too if you need to; this gives even worse performance.

http://www.beyond3d.com/forum/showthread.php?t=26942&highlight=zone+rendering

http://www.beyond3d.com/forum/showpost.php?p=661569&postcount=8

http://www.beyond3d.com/forum/showthread.php?t=16824&highlight=zone+rendering

ftp://download.intel.com/design/chipsets/applnots/25352501.pdf

 

[darkblu]

http://www.beyond3d.com/forum/showthread.php?t=26942&highlight=zone+rendering

http://www.beyond3d.com/forum/showpost.php?p=661569&postcount=8

http://www.beyond3d.com/forum/showthread.php?t=16824&highlight=zone+rendering

ftp://download.intel.com/design/chipsets/applnots/25352501.pdf

you really did not have to. nothing unseen in there. but ok, let's take the full tour.

first, a deferred renderer is called 'deferred' because of the way it handles its primitive commands in time, i.e. it scene-captures. HSR does not participate anywhere in there. rather, it comes as a natural extension, a bonus, and the PVR design takes good advantage of it. intel's zone rendering nevertheless is just as 'deferred' as PVR is by scene-capturing. and believe me, it is nothing like the voodoo1, as you try to jest in that thread you quote.

second, gma900/950 employs ZRT3 (whitepaper), and believe it or not, it does have early-out z.

bottomline being, though ZRT3 may not be PVR, it still is a deferred renderer by any narrow interpretation of the definition.

 

[arjan de lumens]

second, gma900/950 employs ZRT3 (whitepaper), and believe it or not, it does have early-out z.

Interesting; the last paragraph on page 11 of that PDF seems to be saying rather clearly that the Intel chip, like all PowerVR chips, in fact does 100% efficient overdraw removal - although they seem to place more emphasis on the resulting texture bandwidth reduction than on the increased effective fillrate.

I am still a bit sceptical that this is actually the case, though; the benchmark results for gma900, as far as I can see, aren't exactly exhibiting the ~3x effective-fillrate boost that we have traditionally seen with PowerVR chips.

 

[darkblu]

Interesting; the last paragraph on page 11 of that PDF seems to be saying rather clearly that the Intel chip, like all PowerVR chips, in fact does 100% efficient overdraw removal - although they seem to place more emphasis on the resulting texture bandwidth reduction than on the increased effective fillrate.

i believe that's because bandwidth is their major concern by far; the part is severely bandwidth-starved. as i mentioned, if it was an IMR it'd have trouble updating its framebuffer at its full fillrate (4 pixels/clock)

I am still a bit sceptical that this is actually the case, though; the benchmark results for gma900, as far as I can see, aren't exactly exhibiting the ~3x effective-fillrate boost that we have traditionally seen with PowerVR chips.

*shrug* one can always wait and see what changes in their next drivers release.

 

[Ailuros]

darkblu,

If it doesn't increase efficiency as I'd expect it to, then my former bad tasted joke wasn't a joke at all. And before I even believe any fill-rate/bandwidth efficiency results I'd prefer to first see at least on a screenshot how exactly it's filtering.

 

[Arun]

I'd suspect gma900/950 to have early-out-Z and to tile everything, but without Z-ordering, so that you don't have anymore overdraw removal than on a NVIDIA or ATI IMR. In fact, their tiling scheme makes me think of ATI's tbh, which also tiles stuff to reduce that kind of cost.
Is it an OK bandwidth optimization? Yeah, sure. Is it a worthy chip for that? No, it still sucks, always did and always will.

Uttar

 

[Xmas]

Interesting; the last paragraph on page 11 of that PDF seems to be saying rather clearly that the Intel chip, like all PowerVR chips, in fact does 100% efficient overdraw removal - although they seem to place more emphasis on the resulting texture bandwidth reduction than on the increased effective fillrate.

I am still a bit sceptical that this is actually the case, though; the benchmark results for gma900, as far as I can see, aren't exactly exhibiting the ~3x effective-fillrate boost that we have traditionally seen with PowerVR chips.

It seems like Uttar is right and they don't have any kind of Z-first pass (which isn't exactly "Z ordering"). Or if they do, they don't mention it.
Also, if their Z rate is only as low as the pixel rate (and no per-tile min/max scheme kicks in), they're wasting a lot of cycles on invisible pixels.

And Voodoo1 was a scanline renderer.

 

[Fox5]

You can't edit or do some other stuff until you have 1000 posts or so, read it in the appropriate forum:

http://www.beyond3d.com/forum/showthread.php?t=23045

As for PowerVR, I think this was discussed in at least five topics already

EDIT: Rootax, how do you come up with that? AFAIK series 5 was never actually produced in any form, SGX should have some similarities, but it isn't series 5 IMHO.

I thought it was bump mapping and pixel shading that TBDRs weren't suited for. Those chew up fillrate fast, and I don't think the TBDRs can do anything to make them more efficient?

 

[BRiT]

I thought it was bump mapping and pixel shading that TBDRs weren't suited for. Those chew up fillrate fast, and I don't think the TBDRs can do anything to make them more efficient?

Nope. TBDR can utilize defered shading to speed up bump mapping and pixel shading. They dont bump or shade pixels that won't be visible.