[Dalton Sleeper]

Typing error again? GTS on first page, then they wrote GTX 640...

[Geeforcer]

Originally Posted by Fornowagain

Ok PNG lossless versions.

8800GTX vs 2900XT




All four frames in a zip file here.

http://www.sharebigfile.com/file/168282/pics-zip.html

[Arnold Beckenbauer]

Vs:
http://directupload.com/files/myjzmivzgmyw5iwzdgxi.jpg

Can this "fat"first SP do an instruction and SF per clock (parallel) or an instruction or SF?

Why nVidia did forget to mention, that G80 can do one SF every four clocks?

[ChrisRay]

I'll wait for more comprehensive reviews before clucking about IQ (though ATI does seem to have blur in full effect in some modes), but how did NV's PR dept get sshots of R600 in action before NDA lift? I know it's all a grey area, and they probably got a card as soon as anyone else simply b/c they must have better contacts than any of us enthusiasts/reviewers, but it seems a little weird that they're presenting sshots. (Weirder still is that they obviously ran some fillrate tests and their presenting that didn't make me blink an eye.)

And did Google cache Tt's review, ffs?

You dont really think a company like Nvidia wouldnt have gotten their hands on an R600 by this point? Some are already leaking into the retail channel.

[Frank]

Originally Posted by Pete

I can't remember if I've posted this already, but is R600 another case of ATI needing time to fully optimize its memory controller (as it seems to have taken a step beyond the R500 series' one)? Are these app-specific optimizations, with the MC tuned for specific games or engines, or just generically for various AA loads? Or is this just a cover for app-specific optimizations (not specific to the MC)? It does seem like they've got some tweaking to do, given how non-AA framerates are decent but drop like a stone with AA in some games.

That NV comparo that The Inq published claims R600 can filter INT16 at full speed but FP16 at half. Don't they require the same amount of info, both being 16bit? I mean, those fillrate #s NV published look crazy. Then again, R600XT spots G80GTS double the fillrate but benches even with it. So, WTF? Don't tell me NV's "wasting" die space with excess filtering power, b/c it seems a lot easier to tout "eleventy billion shaders!" than "eleventy billion bilinear filtered floating point pixels per clock, kids!"--IOW, that can't be for marketing alone.

Ignoring cache (but continuing with my ignorance), how much bandwidth would an FP16 bilinear filtered pixel require per clock? Is it the same as a single FP32 unfiltered pixel? Or do we have to go one step further and consider the minimum bytes per RAM chip access (e.g., a single FP32 fetch might be more efficient than two FP16 ones)? I'd like to know how each card's theoretical fetching/filtering ability matches up with its physical bandwidth (i.e., if a GPU has more filter/fetch hardware than its bandwidth can support b/c it assumes a certain level of cache hits).

I'll wait for more comprehensive reviews before clucking about IQ (though ATI does seem to have blur in full effect in some modes), but how did NV's PR dept get sshots of R600 in action before NDA lift? I know it's all a grey area, and they probably got a card as soon as anyone else simply b/c they must have better contacts than any of us enthusiasts/reviewers, but it seems a little weird that they're presenting sshots. (Weirder still is that they obviously ran some fillrate tests and their presenting that didn't make me blink an eye.)

And did Google cache Tt's review, ffs?

Like the previous generations, the R600 is optimized for calculating over texture fetches. Add to that, that it is much easier to schedule stuff for the G80 to keep all ALUs filled, and for the R600 lots of portential scheduling hazards for the ring bus (and memory controller).

If you go DX10, calculate as much of your geometry and lighting as possible, the R600 will most likely do very well, if they get the compiler and resource hazards sorted out. And, it seems that (for now) both of those things run quite a bit below par.

On current generation games that use filtered textures for just about anything, it won't surpass the GTX. On next generation games, that use point sampling (data arrays) for much of the workload, it could. It has rather complex texture units, that are very good in point sampling, but much less so in filtered samples, especially if those are floating point.

[INKster]

Originally Posted by ChrisRay

You dont really think a company like Nvidia wouldnt have gotten their hands on an R600 by this point? Some are already leaking into the retail channel.

If they have the right "contact" inside a shared partner, who knows ?
They may get one even before ATI's lesser partners to play with.

But the other way around is also plausible, considering all the "leaks" before the G80 launch last November.

[Razor1]

Originally Posted by INKster

If they have the right "contact" inside a shared partner, who knows ?
They may get one even before ATI's lesser partners to play with.

But the other way around is also plausible, considering all the "leaks" before the G80 launch last November.

Oh they have had it for some time I'm sure , everything kinda points to that, they knew the performance at the conference call, the leaked emails were from last week.

[silent_guy]

Originally Posted by Frank

Like the previous generations, the R600 is optimized for calculating over texture fetches.
...
If you go DX10, calculate as much of your geometry and lighting as possible, the R600 will most likely do very well, if they get the compiler and resource hazards sorted out. And, it seems that (for now) both of those things run quite a bit below par.

This is indeed the same story we heard when R580 was introduced, which was supposedly more future proof. I just don't understand, though, why they keep on targeting games that don't exists instead of focusing on what needs to be accelerated now? By the time those features become important (and they never really did within the lifetime of R580), faster and better HW will be available anyway.

On current generation games that use filtered textures for just about anything, it won't surpass the GTX. On next generation games, that use point sampling (data arrays) for much of the workload, it could. It has rather complex texture units, that are very good in point sampling, but much less so in filtered samples, especially if those are floating point.

I understand that it's possible to shift some of the texture filtering load to the shaders and how that would benefit R600. But I don't understand the incentive of a developer to do this (and would very much appreciate it if they provide their thoughts on this.)

If it is currently doable to use the bandwidth of the texture units to do all kinds of fancy stuff that would otherwise required shader calculation power, doesn't it also make sense to keep this functionality in the TU's and use increased shader power for other stuff (like GS)? It seems that memory bandwidth is about to increase significantly during the coming years, this stuff has to be used for something: if ROP usage isn't increasing much, isn't it likely that it can be most efficiently used for texture operations?

I went through the slides of the Cascade demo: It seemed to me that they were still relying a lot on clever texturing techniques in combination with GS to get the best result. (One again, some insight from a 3D programmer would be most helpful.)

Also, even if there will be a shift, isn't this something that will take multiple years to complete? Most engines will still need to run decently on DX9, so for the foreseeable future, DX10 usage will be primarily used to add nice effect here and there. And there's, of course, the simple fact that programmers can fall back on tons of well documented existing techniques to make certain things happen, while they'll need to learn a bunch of new stuff to make full use of DX10.

Edit: If you saw the human head demo that was recently posted on the Nvidia developers blog, I had the impression that it used an incredible amount of textures to get just the right light behavior. I'd be surprised if this kinds of results can be obtained as efficiently with shaders.

[Subtlesnake]

Originally Posted by Rangers

I was actually somewhat impressed with the benches on this card, it seems to beat a GTS and approach a GTX in many cases. For $399 it seemed worthy.

However, once you factor in the lame IQ and high power requirments, it's probably a no sale.

BTW, I like how the Vr-Zone article several times blames poor results on X2900XT's "bandwidth", when it has more bandwidth than any other card. Clearly they mean texture capabilities, but apparantly dont understand it. I expect to see most sites not understanding the extremely low texture abilities of R600 are clearly responsible for it's poor performance.

Well, then why is there such performance variance from application to application? Wouldn't you expect texturing demands to be more consistent across games?

Given the huge performance increases we've seen over the course of a couple of driver releases and the fact that in certain instances the X2900 is close to the 8800 GTX (while in others it performs at X1950 XTX levels!) I think we can point to other factors.

Do we even know how R600's texturing capabilities stack up to R580's?

Too bad for ATI R600 contains over 700 million transistors, and engaging in mid-range sub $300 price wars is not where they really want to be with the product.

Isn't the 2600 intended for the sub $300 price points?

[Frank]

Creating geometry and lots of shader power are good things, as long as most of your potential game buyers have them. Creating geometry is needed for physics, clothing, hair, fluids and deformable terrain, while lots of shading power are great for weather and dynamic lighting.

[silent_guy]

Originally Posted by Frank

Creating geometry and lots of shader power are good things, as long as most of your potential game buyers have them. Creating geometry is needed for physics, clothing, hair, fluids and deformable terrain, while lots of shading power are great for weather and dynamic lighting.

But does it replace the need for texturing throughput in any way?

[Razor1]

It won't replace texturing through put, but if there is a shift in the bottlenecks, which it seems probably won't happen at least not to the degree the r600 is held back (if infact it ends up the texture units are what is holding it back) then the extra shader power and again possible GS performance advantage will come in. I say possible because the GS problem seems to be actually something trival, but again take this last part with some salt I'm not certain if its really trivial or not but thats along the gist of what I hear.

[Frank]

Hm. What textures can we remove?

Simple, colored ones? Perhaps, for materials like wood, leaves, rocks, plastic etc. As long as we make sure each polygon uses only a single material. But that won't work for details and most objects.

Normal maps and bump maps? That could be done, in many cases, as long as the surface is reasonably smooth, or with lots of straight lines. But probably not for (N)PCs and most objects. And I don't think artists would like that.

Light maps? As long as you only use a single material for each polygon, you could remove most of them.

So, you could theoretically get rid of many textures that are used for static terrain.

[Geeforcer]

Originally Posted by Geeforcer

Sorry for reply to myself, but after looking at this picture again, it seems that one is blurrier then the other. Not a lot, nothing compared to the jpeged mess - but noticeable. Of course, there is a chance that after staring at the monitor for hours my eyes are playing tricks, so what does everyone else think?

[nutball]

Originally Posted by Geeforcer

Sorry for reply to myself, but after looking at this picture again, it seems that one is blurrier then the other. Not a lot, nothing compared to the jpeged mess - but noticeable. Of course, there is a chance that after staring at the monitor for hours my eyes are playing tricks, so what does everyone else think?

I think that your question is an answer in itself. After hours of staring you're starting to doubt your own eyes.

[leoneazzurro]

Originally Posted by Razor1

It won't replace texturing through put, but if there is a shift in the bottlenecks, which it seems probably won't happen at least not to the degree the r600 is held back (if infact it ends up the texture units are what is holding it back) then the extra shader power and again possible GS performance advantage will come in. I say possible because the GS problem seems to be actually something trival, but again take this last part with some salt I'm not certain if its really trivial or not but thats along the gist of what I hear.

I don't think it's texturing that holds back R600. Its filtering capabilities, from what comes from the slides, are far beyond R580, and while I think them being inferior to G80, I don't see so many games being only and so texture limited (and anyway in the worst case texturing power should be very similar to GTS according to what R600 is supposed to be). Other very strange topic of R600 is its behaviour with cranking up AA and resolution. Being a bandwidth monster, it should perform better where bandwidth is needed but it seems it lose a lot of performance (even in comparison to GTS) where it should have a lower hit.

So I think there are a lot of problems in optimizing with this architecture. In the VR-Zone review I saw a great result with 3D mark 06 sinthetic vertex shader, where pixel shader test was simply underwhelming. Probably the co-issue thing needs still a lot of work in order to extract maximum performance...

[Jawed]

Originally Posted by leoneazzurro

Probably the co-issue thing needs still a lot of work in order to extract maximum performance...

Perlin noise test contradicts that assertion.

Jawed

[BlizzardOne]

Originally Posted by Geeforcer

Sorry for reply to myself, but after looking at this picture again, it seems that one is blurrier then the other. Not a lot, nothing compared to the jpeged mess - but noticeable. Of course, there is a chance that after staring at the monitor for hours my eyes are playing tricks, so what does everyone else think?

I took the original PNG's and used ATI Compressonator to do a comparison, put a negative filter on them to make the differences a bit easier to spot, unfortunately they're close to 4mb each, thus too big for imageshack! If anyone can suggest an alternative free host, i'll throw them up there..

There are differences, but in motion, i think 99.9% of them will go unnoticed..

[Geeforcer]

Originally Posted by nutball

I think that your question is an answer in itself. After hours of staring you're starting to doubt your own eyes.

So, is that a "yes" or a "no" to one picture being blurrier?

[Pressure]

Originally Posted by Razor1

Oh they have had it for some time I'm sure , everything kinda points to that, they knew the performance at the conference call, the leaked emails were from last week.

I wonder because the Geforce 8800 Ultra has "Panic Button" written all over it.

[leoneazzurro]

Originally Posted by Jawed

Perlin noise test contradicts that assertion.

Jawed

Then why "pixel shader" test was so bad? This GPU should be way off the GTS in pixel shading power, even counting the "missing MUL" on the G80's based GPU.
And in the 3Dmark sinthetic test is on par with GTS. So? Where's the problem IYO?

[Geeforcer]

Originally Posted by BlizzardOne

I took the original PNG's and used ATI Compressonator to do a comparison, put a negative filter on them to make the differences a bit easier to spot, unfortunately they're close to 4mb each, thus too big for imageshack! If anyone can suggest an alternative free host, i'll throw them up there..

There are differences, but in motion, i think 99.9% of them will go unnoticed..

I had the same problem with PNG. The original version of that picture was a 3.4mb file. I ended up going with 97% jpeg in Infran. However, since the quality of source material and the save settings are the same for both shots, it's not that one would suddenly start looking worse then the other.

[nutball]

Originally Posted by Geeforcer

So, is that a "yes" or a "no" to one picture being blurrier?

It's a ZOMG my eyes hurt from the staring!

Seriously though, I think the one on the right is blurrier around the white highlight, the one on the left is blurrier below that in the "valley" in the foreground. But really I think we'd need to see them in motion to assess the real impact in IQ (and honestly I think that a short movie clip would probably show a thousand times more info than a single still).

[Razor1]

Originally Posted by Pressure

I wonder because the Geforce 8800 Ultra has "Panic Button" written all over it.

Actually not really a panic button, look at the price of the 8800 Ultra, that kinda started the hints off. I guess it was there just incase, but with a price like that, I think nV already had a good idea of what is coming out, thats was 2 weeks ago. But those emails also have the pics of the different AA modes. So either they go someone to take the shots for them or they took those shots themselves, either way, they had access.

[Geeforcer]

Regarding Ultra, I wonder what is being talked about here.