The Official X1900 Reviews Thread

Lets look at the 7900, rumored to be 32 pipes and 700-750 mhz.

Of course these are just rumors, etc etc.


Lets say it's 700. Will probably be higher, but lets say 700.

That's 27% faster than 512 if nothing changes.

But wait, now add that it's 32 pipes, and every pipe is 27% faster. You end up with it behaving like a 40 pipe 512. Or around 70% faster.


And it seems like, Nvidia actually SCALES.

These numbers should be scary for ATI. If it's truly 70% faster they will only be able to compete in one game, FEAR, and that just barely..

This may totally not pan out. 7900 may only be a little faster too. But that's a scenario.

 

Not too mention Nvidia already wins all the openGL games, many older games, and many of the benchmarks with no AA and AF, too boot. They do that right now. They dont need to wait for 7900 to do that.

 

There is no evidence showing the pipes being a magical 27% faster. I wouldn't be so quick to assume a 550 to 700+ MHz increase along with an additional two quads would be achievable going from 110nm to 90 nm either. I've already said this many times.

 

Hmm, you're right. I don't know quite what I was thinking about.

 

I think you made a mistake in your post. You assume that 27% faster clock speed == 27% more performance. That's not the case. GPUs (and CPUs) do not scale linear with increased clockspeeds and added pipes. Additionally at least the first bunch of G71 cards will use GDDR3 memory and that memory will probably not be any faster than the GTX512's memory. So the theoretical 70% increase will only result in a 10%-30% increase in performance (probably depending how memory bandwidth starved the G71 will be).

 

Yeah but who really cares about non-AA benches . No AA on highend cards and ultra highend cards is a waste of time and is purely academic . I wish people would stop even mentioning it .

 

Non-AA benches are useful when you're trying to determine if a game is really as ALU-bound as it's said to be.

Also remember that current NVidia hardware sort of has a 2:1 ALU:TEX ratio - so ATI's "advantage" isn't as clear-cut as it first appears in supposed current, bleeding edge shader-heavy titles like FEAR.

R5xx in FEAR isn't actually benefitting directly from its excess of ALUs, so much as its benefitting from its ability to maximise the utilisation of ALU and texture units, so that they spend as little time as possible idle. It's a symbiotic relationship, generally speaking.

Jawed

 

what kind of gains does Toyshop see going from R520 to R580 (assuming I skipped past any mention of it)

 

Double, supposedly.

Jawed

 

I've heard somewhere(don't remember) that it's about twice as fast on the R580 than it is on the R520. Don't quote me on that, I just remember reading that somewhwere.

 

Precisely ............. academic .

It doesn't matter , or at least it shouldn't matter to gamers in the particular title , since they should, I hope, be playing with AA and AF enabled . Trying to figure out if something is future proof or efficient or does what it is supposed to or is pipelin/alu/tex/mem bandwidth balanced can be helped with non AA tests . But ultimately it is an academic pursuit .

 

I don't think Nvidia will be increasing the TMUs that much though. Even with the G70 they have 24 pixel shaders but only 16 TMUs. If they go for 32 pixel shaders, perhaps they'll go for 24 TMUs. I see no chance of a 32 TMU chip. It would need massive amounts of memory bandwidth for that to be at all effective. Isn't that why ATI stuck with 16, because they don't have the bandwidth to adequately feed 24 or 32 TMUs?
________
CRESSIDA

 

Uhh, I think you mean ROPs, not TMUs.

 

But the shaders in current titles have at least a 3:1 ratio of arithmetic to texture operations, so they can't be useless! Look at 3D Mark 2006 for example. The X1600 XT is able to match the 6800 GS in many cases despite having only 4 texture units (compared to 12 in the GS).

Yet, the X1900 XT only performs slightly better than the GTX 512, despite having double the number of 'full' PS ALUs. So I don't think there's a texture limitation at all. The card is probably just being held back by a lack of bandwidth - the memory clocks are virtually unchanged from the X1800 XT, yet there are 3X the number of PS ALUs to feed. The same thing will happen to the 7900 unless it debuts on faster memory.

As for the extra die space taken up being a 'waste' consider that the die has only increased by 20%. Yet performance over the X1800 XT has increased by a good 30% in shader limited titles.

 

First of all. the 512GTX is a vapourware product. The X1900 eclipses the 256GTX product. You can't buy one and the X1900 hasn't even had its first driver update yet. ATI has also indicated they will put more effort into tuning OpenGL and the results are already showing. Considering that lots of OpenGL games are like Doom3 are specifically architected to perform well only on Nvidia hardware.

When games start implementing features like fetch4 and PS3.0 the ATI architecture will truely be able to spread its wings.

Keep in mind that GDDR4 goes into production in March and the X1900 has a memory controller that will support it. GDDR4 is 56 percent faster.

The real question here is why does it beat a chip with 50% more pipelines and more memory bandwith.

With a 256-bit bus Nvidia can make 8 64-bit memory accesses per memory clock. ATI can make 16 32-bit memory accesses per memory clock. If ATI can saturate the bus with the existing TMUs then texture cache because more important then number of TMUs.

 

It's not a vaporware product. It does have limited availability, but you can indeed purchase one, if you're willing to wait (how much depends upon when you try to buy...looks like usually a couple of days to a week).

Fetch4 is really minor compared to some of the other things the R5xx supports, in terms of how it will do against the NV4x parts.

What, and you think the G71 won't?

And fewer ALU units and lower clock speeds.

Well, they can't in the case of compressed textures, that's for sure. Even with zero texture cache hits, compressed textures are only 4 bits per texel, or 16 bits per texture access. At 16 bilinear filters per clock, that's 256 bits per clock. So with zero cache hits, compressed textures are only taking up about half of the memory bandwidth (DDR memory clocks are typically somewhere near twice the core clockspeed).

Now, if you are running with trilinear filtering enabled, you are basically guaranteed to have at least a cache hit efficiency of 5/8, so the above 256 bits per clock drops to 160 bits per clock.

Then, if you're really good at texture cache hits, and since texture filtering is designed to average one texel per pixel (without anisotropic), you can improve the above by another factor of four, dropping down to only 40 bits per clock required for texture accesses.

So, if your texture cache is doing its job and you're doing trilinear filtering, you could go up to 64 bits per texel (16-bit FP textures) without saturating the memory bandwidth (this would be 320 bits per core clock...remember that due to the higher clock of the memory controller the bandwidth is roughly 512 bits per core clock for current architectures).

 

The bit-tech's Radeon X1900-series roundup: http://www.bit-tech.net/hardware/2006/02/10/radeon_x1900_series_roundup/1.html
There is something in it that got my attention. In http://www.bit-tech.net/hardware/2006/02/10/radeon_x1900_series_roundup/6.html Tim Smalley remarks that the X1900XTX in F.E.A.R. has an advantage in best playable settings having the posibility to use 2x performance adaptive antialiasing over the X1900XT which can only use 2x normal antialiasing to obtain smooth framerates at the settings he uses. Thinking about how performance adaptive behaves, and that in this mode you only obtain half the texture samples than geometry samples, half of two samples is only one, so I believe that in theory you shouldn't obtain any alpha texture antialiasing and there shouldn't be any difference in image quality/performance between 2xPA AA and 2xAA. Looking at the samples with DX9FSAAViewer I obtain this sample positions:


I said this to the bit-tech crew in http://forums.bit-tech.net/showthread.php?p=1176074 but Tim said "I'm pretty sure that there's a difference in performance and quality between 2xAA and 2xPA AA" I don't want to discredit Tim in any way, he does great reviews that I love to read, but I'm not sure if he's accurate here and I want to know if there is really any difference in image quality/performance between 2xPA AA and 2xAA. Can someone help me?

 

http://www.tech-hounds.com/review17/ReviewsPage1.html

Somewhat different reviewing style which could do with some design input (for clarity's sake). What's notable are:

• comments on IQ based on moving images, not just screenshots
• an analysis of performance in FPS "bands" over the duration of a benchmark run

Jawed

 

Does nVidia use SS 1,5x1,5 or something similar on SW Kotor? The blurring and poor AA efect on near-vertical lines is just horrible (GF7 screenshot / R580 screenshot)