The G92 Architecture Rumours & Speculation Thread
- Thread starter Arun
- Start date
- Status
IIRC
TAU is essentially fetching the textures.
TFUs allows for a smaller performance hit when enabling AF, as it can filter the units faster than new ones are being fetched?
I'm not quite sure either. Whenever I try to go through G80's architecture in my mind I always think of this picture (CO Arun):
http://www.beyond3d.com/images/reviews/g80-arch/TheG8.png
and realize it's far too much for me to take in sometimes when broken down into the finest pieces.
Dammit, I miss ROPs/PixelS/VertexS/TMU being essentially the key ingredients. Now every thing's all decoupled and minced into finer units, and you're not the only one that finds it confusing sometimes, even if you'd like to understand.
Blazkowicz
Legend
damn I really need to get one of these GPU with "The Fucking Unit"!
It's ironic that some complain there is no info prior to a rumored hard launch and others would complain if there was a soft launch. I guess companies can't win.
I linked it in one of my previous posts, as I said already. It's in Rys' first G80 article. Check the "Reviews" area of this site.
OK, I've spent too much time hacking out an explanation of TAs and TFs, so I'll post it hidden behind spoiler tags if you're so desperate that you'll risk bad info. Otherwise, either read Rys' article or hit up Google or just wait for someone knowledgeable to explain.
At the ever-present risk of talking out of my ass, I'll take a stab at summarizing. The TA fetches the data from memory (color, in the case of a texture). The TF performs a bilinear filter on it per clock. Since trilinear filtering requires two bilinear filters, and the G80 has two TFs to each TA, the 8800GTS/GTX essentially offers free trilinear filtering in the sense that that second TF would otherwise have nothing to do if only bilinear were called for, as the single TA can't feed both TFs simultaneously. I think the idea is the TA fetches the data for TF1 on the first clock, then for TF2 on the next clock while TF1 performs the second bilerp to complete the trilinear filter. I'm losing my grasp on things, though, so I'm not exactly sure how that second TF offers free 2xAF, because it seems you need as much data fetches as you need filter ops. I can't even figure out how my example works for trilinear filtering, as it seems each TF would require data from a different MIP map each clock, unless texture cache is involved somehow. So, don't trust any of this.
ok i think i just had a light bulb moment, i know this is basic stuff for you guys but here it goes. the G80 has 8 shader processors, each of which have 4 TAU and 8 TFU. essentially TMU = TAU. the 36.8 GT/sec fiilrate often used fo the G80 is INTERNAL throughput, correct? Based on 575mhz core x 64 TFU. This helps decrease the impact of filtering such as AF because the TA units dont have to wait for TFU, since there is twice as many TFU and thus data is always ready for the TAU? But due to fact there are a total of 32 TAU, actual external texel throughput on the G80 is 18.4 GT/sec. And it is able to reach very close to this theoretical limit because of adequate memory bandwidth....unlike what appears to be going on with the 8800 GT. Unlike the G80 it only has 7 shader processors, but like the G80 has 8 TFU per shader processor. unlike the G80, its number of TFU = TAU, thus is has 56 TFU (7x8) as well as 56 TAU (7x8) [G80 has 32, 8x4]. However, there is a bottleneck present in the 8800 GT that appears to prevent it from even getting close to its theoretical 33.6 GT/sex fillrate. instead it appears to reach only around 25 GT/sec. and so far people think this inefficiency is due to either memory bandwidth, speed of shader processors or perhaps both. I myself can understand how memory bandwidth could play into this, but I cant even start to see what speed of shaders would have to do with it???
now i also need to understand how ROPs play into this.
now i also need to understand how ROPs play into this.
Last edited by a moderator:
2600 X2?
Not even close.
I believed that nvidia hidding something in G92 they just not active it yet.
Mid-November:
"Today nVidia announced all GeForce 8 Series graphic-cards will support D3D10.1 through a later driver-release, which will come contemporaneously with Vista SP1.
They decided to notify this relative late, because in their opinion they have more to offer than just a filled feature-list."
Yes, a beast in power-consumption and BOM, since it should be delivered with 1GiB fast GDDR4, which means for each GPU 1GiB -> 2GiB, because of AFR...
Mid-November:
"Today nVidia announced all GeForce 8 Series graphic-cards will support D3D10.1 through a later driver-release, which will come contemporaneously with Vista SP1.
They decided to notify this relative late, because in their opinion they have more to offer than just a filled feature-list."
Hope fully.We can see it more clearly when GeForce 9 series release.Hope its based on G9x or something and we might lucky to see G92 rebranded as GeFore 9600 or even 9700.
Mid-November:
"Today nVidia announced all GeForce 8 Series graphic-cards will support D3D10.1 through a later driver-release, which will come contemporaneously with Vista SP1.
They decided to notify this relative late, because in their opinion they have more to offer than just a filled feature-list."
That whould be a litle stupid because they are losing lot of market just because they don´t have DX_10.1 suport/sticker in the box.
So take your conclusions. Things should be very clear since day one, not like that way.
Look at the graphs!
Pretty common though, everyone does that (yes, even daamit). The PR dudes know how to trick the buyers.
Spend a day downtown and ask this Q: Do you know what RV670 is? If you ask 300 people I bet you'll get 299 incorrect answers.
- Status
Similar threads
