The New and Improved "G80 Rumours Thread" *DailyTech specs at #802*

I do. And I did, above. You are quite right - 'a bit under 2' is a passable estimate for going from bilinear 1X to trilinear 16X AF, on, say, an FPS type scene.

 

I'm not really clear on how drivers need to be optimized so that the hardware can correctly auto-balance itself. I don't want to jump to any conclusions on the basis of badly translated, pre-released reviews, but, something doesn't sound right there.... If fixed allocation works better than dynamic, then, their balancing algorithm missed something important, or paid attention to something it shouldn't have. I hope the balancing algorithm isn't hard-coded.

 

Good, other trustworthy stores in Spain say that there will be availability of the GTX

 

D3 engine uses a lot of texturing instructions, with little (or none?) independent ALU instructions - the Humus tweak investigations showed wildly varying performance depending on how AF was applied (control panel or application), etc. along with a performance benefit in replacing one texture operation with math, so texturing is a significant part of the performance of D3 (otherwise, why did B3D use D3 in ultra quality as the basis of an AF scaling test?).

Jawed

 

That's quite odd, since right here "next door" there are stores with plenty of stock of both G80 variants already.

 

I think that would mean its not odd

 

Sorry, i've just realized he was quoting himself.

 

Jawed's point is that R520's TMUs are bound by the relative paucity of ALUs, so they may sit idle some of the time. R580 alleviates this by tripling the ALUs. R520's weak point is lack of ALUs and R580's, lack of TMUs, both of which may lead to underutilisation of the other.

I think he means it as better die usage efficiency, meaning they're in use almost all the time? I dunno.

Edit: trini, where's the 7900 beating up on the GTS? I see it at worst tying in that one review (which we shouldn't even be discussing if NV's NDA had any weight or was at all fair, but that's for another thread), but if G80's TCPs are really being artificially limited to specific tasks (PS/VS), maybe the GTS' one fewer TCP leads to a less ideal hard-locked PS:VS ratio than the GTX. Plus, drivers--it's so early!

It's also interesting to think of G80 "losing" some of its impressive # of TMUs in DX9 games when a TCP is processing vertices. I mean, current DX9 games don't use the TMU much with vertex shaders, right?

 

It's also worth noting that if G80, in these benches, is running with a static allocation of SPs to VS and PS work, then the associated TMUs will also be tied-up in the same ratio (assuming that TMUs are dedicated to their respective SPs).

So, if 32 SPs are allocated to VS work, the remaining 96 are allocated to PS. I guess that means PS has access to 48 TMUs (24 TAs + 24+24 TFs). These DX9 games will therefore presumably "waste" 16 TMUs, with such a ratio.

Of course this ratio may vary from one game to another. And who knows, the ratio may suit GTX more than GTS. Additionally, the ratio may suit non-AA/non-AF over AA/AF modes.

Who knows?...

(Wonders if the VS-specific TMUs are actually idle - perhaps they are employed even when there is no vertex texturing.)

Jawed

 

Aren't you all excited? I know I am. Can't wait till 15:00. I just worked my way through some documents (they were laying in my mailbox for a few days, but I forgot to check it ). The documents are pretty big. For instance the GeForce8800 GPU Architecture document alone is 53 pages long and the DX10Technical Briefing is 24 pages long.

Important things... Shimmering effects are gone on G80 and it has better IQ than ATIs current high end offering.

I could post some stuff here already but I haven't checked if it was already posted here... and I'm not bound to an NDA.

Some interesting stuff and buzzwords you'll see tomorrow:

- Purevideo HD
- XHD Gaming (of course the 'G80 is built for XHD')
- nVidia began designing the G80 in the summer of 2002 and it looks like it was unified from the start

 

I hope this isn't another Purevideo type problem. Didn't someone somewhere say it was "DX9 compatability"? Does this mean it will only load balance under Vista/DX10?

 

Take no chances.

Scan it, and post the pdf asap!

 

CJ - since you're not bound by an NDA, could you say whether or not a TCP is bound to work on one thread type at a time, or whether a single TCP can process both vertices/pixels simultaneously? If it is just one thread type at a time, can you say anything about how often the TCP assignments are changed?

 

So, it is confirmed.
When David Kirk was interviewed by bit-tech last year he was actively spreading FUD.

 

I could post the PDF, but I'm not gonna risk exposing my sources.

Here's some from the beginning:

and some more marketing stuff:

This is also interesting as I always thought that ATI's R600 was Microsofts reference DX10 card:

One sec and I'll try to get some info for the tech-freaks in here to end all of the wild guesses about the way G80 works.

 

Well, Microsoft's reference for Dx10 development != Induvidual developer's reference for DX10.

 

Thanks Pete but it seemed to me that they were comparing synthetic texturing numbers at one point - which would not be impacted by ALU power.

That same pconline review showed the 7900GTX outpacing the 8800GTS in several games. Check the tables, not just the bar graph.

Yeah, interesting observation. Hopefully it will be a lot clearer tomorrow.

 

Microsoft chose G80 as the reference DX10 GPU ?

But, but...

 

Oh dear, but at least we have "FP16 bilinear texture filtering is also performed at 32 pixels per clock".

Jawed