512MB GeForce 7800 GTX

Xtpe?

 

DX9 doesn't specify performance. Anyone developing a DX9 engine in the last 18 months is going to stick with shaders that won't bring an R300 to its knees. Also, more complex shading doesn't neccessary mean longer shaders per pixel written, but could mean more fillrate, as the scene is rendered from multiple angles.

Well, most titles are targeted for the xmas season, but this is orthogonal to the question of whether those titles will feature massivelly long shaders that favor the R580. I don't buy the argument that ATI designed the R580 for titles coming out this xmas. IHVs have been playing this workload game for a long time, and by the time their HW ships, the workloads have changed. When the NV3x was being engineered, it looked like the future was stencil shadows. We had D3 videos, we had 3DM05. People thought this was going to dominate workloads. Turns out it didn't.

Now we have the idea that long shaders will dominate workloads. But because of shader model limitations, there is only so much you can do in the shader. Many algorithms require auxillary buffers to be rendered, and that lowers the shaderop/pixel ratio.

I don't know what the future workloads will be, but I don't buy the talking points that try to position ATI as an oracle that always "times the market" with their parts when they are needed.

 

Nope, Tarinder wrote that article before the news of some increased X1800 clocks came in. He's talking about what we're talking about in this thread, or so I imagine

 

One of few tests i see them overclock the ATi card.

 

Einstein you are not! ....

 

Nope, that's not the right way to look at it at all in my view. Not when you see a game like FEAR, which has been in development for longer.

Games developers should be making eye-candy options - e.g. the quality of reflections, or the use of soft shadows (I'm not saying that FEAR is an exemplar of good-looking eye-candy in these respects). The game should demand a minimum spec (e.g. DX8) and then provide enough eye-candy to last a good while beyond the initial release of the game.

If R580's anything like RV530, then the texturing performance it'll deliver will be way beyond what it appears to have, theoretically. RV530's four texture pipes:

http://www.beyond3d.com/misc/chipcomp/?view=chipdetails&id=104&orderby=release_date&order=Order&cname=

easily keep up with NV43's 8 texture pipes.

http://www.beyond3d.com/misc/chipcomp/?view=chipdetails&id=71&orderby=release_date&order=Order&cname=

So R580 isn't just about shader arithmetic. The one thing R580 is plainly not going to be getting any advantage from, for some time yet, is dynamic branching.

The 1.1ns memory that we're expecting to see in the 7800U tomorrow is also the same memory we're expecting to see on the top-end R580. Not to mention the more efficient memory architecture.

So R580 (if it had been released for Christmas) would have covered all the bases available to high-end games:

• lots of bandwidth
• lots of arithmetic
• lots of texturing

I don't think there's anything left. A high-end game would (like HL-2 and D3 last Christmas) sell lots of the latest high-end (as well as mid-range) graphics cards.

Ironically NV3x is no better at stencil shadowing than the competing R3xx cards. If D3 had been released a year earlier, NVidia would have lost even more sales. Lucky for NVidia that D3 didn't release until after NV40 was released.

FEAR and Chronicles of Riddick both use stencil shadowing, as does Q4. Games like HL-2 get the shit ripped out of them for their arse shadowing.

Obviously we have Quake Wars and Prey both coming with stencil shadowing. I wouldn't be surprised to see another year's worth of games with it.

How? If you're rendering to an intermediate result, those operations still count to the final number of ops required to put a final pixel together. Even if the render target is 1/4 screen res. That work can only be done by a shader.

R580 looks like it will be a good all-rounder.

Whereas R520 looks like a "gentle introduction" to ATI's new out-of-order/ring-bus GPU architecture.

I'm still somewhat stunned that R520 is showing such desultory speed-ups in many places (I keep moaning about the seeming lack of any evidence for the out-of-order scheduler's efficiency gains) - but then you look at something like FEAR and see that R520 is 80%+ faster than R420. Not sure how much of that is coming from the pixel shaders though - but some of it, surely?...

---

I will admit FEAR is a bit problematic as a bastion of current high-end games, because in many ways it doesn't look like it should be bringing GPUs to their knees. Is it sloppy code? Dunno.

Jawed

 

I didn’t say it did. What DX9 (the whole software and hardware environment) allowed was more ALU utilisation than previous generations.

It doesn’t have to bring it to its knees. But the simple fact is that ALU utilisation is increasing, and has done so over the course of the last few years.

I’m not saying they will favour R580 massively or exactly, but any increase of ALU utilisation will favour R580 in relation to R520. The simple fact of the matter is that NVIDIA’s ALU capabilities have been scaling (both NV40 from NV30 and G70 from NV40) which is another factor that will favour R580 in relation to R520.

As a note, the organisation of the pipeline means that any operation that spends more time in the ALU than it does getting texture information back or writing a pixel is likely to be a gain for R580 in relation to R520. At this point in time I wouldn’t expect the gains, clock for clock, to be terrifically large, but I would expect the gains to be larger in more recent titles.

I’m not saying that either, I’m saying that concentrating on IQ performance first and then going back an increasing the ALU capabilities would look a far better move had they been here at their correct junctures. Should R580 turn out the way that its expected then there are far more practical reasons for doing it this way in that in all likelihood they would get their SM3.0 design out the way first with R520, then go back to rework the “quad” to increase the ALU power with a relatively simple part (RV530), then replicate those quads for the high end later.

 

http://www.monarchcomputer.com/Merchant2/merchant.mv?Screen=PROD&Store_Code=M&Product_Code=190348

 

What are the current GTXs clocked at?

EDITED BITS: Oh, and frontpaged...thanks.

 

430/430/470

Seems the 512MB is clocked at 550/550/550 according to Sampsa, would be interesting to know why nvidia have gone back to all domains being the same. Maybe their funky heatsink can handle it and with the standard GTX they played safe and just had the vertex domain increased ?

Maybe the reviewers can spread some light on that.

In regards to what you should use your transistors for I am heartily glad that they are now being used for better effects. There's been too much time and effort gone into getting the nth degree of AA and AF in games for me, especially AA. Though I do agree that a lot of people will disagree and I did like transparency AA.

I'd far rather have a creature hiding behind an HDR beam of light than some jaggies where I have to stop and look for them.

 

I'm an AA whore so it's alright with me if they want to put more transistors towards it.

Can ya picture two of these beasts in SLI?

 

With the 1XXX series cards you can both

 

The GT's clock domains are all in lock step with each other.

 

so this thread is getting long. Have you figured out if the 512MB 7800GTX is going to kick butt or not?

 

So much so that some consider it a mugging... a $700 mugging.

 

*thinking aloud*
All I need is a bunch of benchmarks to help me rationalize that price point and I'll be set! The 1800XT 512MB is $600 or $550? Either way at that amount it is insane prices so the differences dont matter as much as performance. With the 1800XT u get the killer IQ but the 7800GTX 512MB is faster and has a nice heatpipe heatsink...that sucker must get HOT!

 

There's text in front of them.

And how sure can we really be that wherever the R520 poses somewhat unpredictable performance it's due to drivers not being yet fine tuned for specific games?

I'd rather buy the more obvious theory that architectures have both advantages and disadvantages and it's not really a given that any change within an architecture will always deliver the gains some would expect.

Fear is a weird case and I still have no idea what is really going on in there. I've been expecting myself games to finally tax the current GPUs more, but what I get on my screen doesn't justify the enormous performance penalty.

 

I don't think it's either/or here and I'm not even suggesting anything specific.

But it's undeniable that with a fairly new architecture (not just in one area, but in several key areas) driver tuning is a new sport. As I said earlier, we have one case where tweaks for older architectures make R520 slower.

Maybe an interview with the Monolith is due.

The XB360 game Condemned (from Monolith) appears to use a very similar engine and with both of these games "finished" (Condemned is a launch title) they should be up for an interview...

Jawed

 

What makes me having second thoughts is that they had working prototypes for a fairly long time. With the exception of the OGL-AA tweak I haven't seen anything yet that suggest immature or untuned drivers; in fact I get the feeling while reading reviews that R5xx has been a lot longer around then just a couple of months.

One case could easily be an exception.

I don't think you'll get out of an interview what I'd really like to read. What I'd like to see is a developer having a closer look with the appropriate tools over it and try to find if and where any bottlenecks could be; sadly most of them don't have enough time to waste for such ventures.

 

In F.E.A.R., without AA, the X1800XL is about twice the speed of the slighly lower clocked 8-pipe X800GT. I don't think any of the speed gains whatsoever come from improved pixel shader performance.

Are you sure R520 is out of order? Just because there are a lot of threads and the dynamic branching is good doesn't mean it's out of order. In fact, it doesn't make much sense, because the shader compilers can do most of the ordering, and they'll be a lot more effective than hardware for pixel shaders due to their predictability. I think OOO is much more important for general CPU code.

I don't think R520's architecture has a leg up on R420 in much more than dynamic branching and dependent texturing due to better latency hiding. I was looking at Shadermark data from two B3D reviews:
http://www.beyond3d.com/previews/nvidia/g70/index.php?p=10
http://www.beyond3d.com/reviews/ati/r520/index.php?p=18
Popping the tables into Excel, I found some interesting things. All the following are comparing performance per pipe per clock, and use geometric averaging:
1. G70 is barely faster than NV40. (1%). Some losses here and there, a big loss in dynamic branching (???), and gains in HDR.
2. In shaders 2-7 (comparatively simple lighting shaders), NV40 has a big lead over R420 primarily due to the free FP16 normalization (as seen here, but note the shaders are a different). In those same shaders, R520 is barely any faster (~5%) than R420.
3. In shaders 8-26, while some shaders still benefit from free FP16 norm, NV40 is 34.6% faster than R420. R520 is 33.4% faster than R520.
4. Overall, G70 is 25% faster than R520 in shaders 2-7, and 3.5% faster in shaders 8-26. On a shader by shader basis, there is quite a bit of variation.

(Aside: Given #2, it's very surprising that the X1800XT can catch the GTX in Doom3, especially when also considering stencil performance.)

Anyway, don't expect miracles from the scheduler. Previous hardware was quite efficient at getting everything it could out of the shaders. The major advantage of the R520 shading architecture is dynamic branching, as well as having the ability to change the ratio of texture units to math shader units. Dependent texturing looks like it's faster also, but to a much lesser degree. In any case, it looks like R520 is about as fast as G70 per pipe per clock unless FP16 normalization makes a big difference. Looking at games with lots of pixel shading, however, it seems like ATI's compiler makes a difference, as the shadermark results aren't translating into reality.

Right now, it does look like ATI spent a lot of transistors for nothing. But as someone said earlier, the 3:1 math to texture unit ratio is when it makes most sense. When R580 comes out, I think it would beat a 32-pipe, equal clock speed 90nm G7x, but only in games as math heavy as FEAR and COD2. I'm sure ATI is counting on XB360 to make this the norm. If dynamic branching becomes a factor, then 580 will naturally blow this hypothetical G7x away, but I doubt we'll see it happen much, if at all, next year.

I think one area ATI might take advantage of its branching capability is by using early out for N dot L lighting, which would be really beneficial with multiple lights. It wouldn't be an easy driver optimization, and may only be possible via hand tuning, but it could definately be worthwhile under the right circumstances.