I dont know exactly but I vaguely remember the use of 0.9ns in the GTX512 reminded where it got really hot.

Oh I'm tempted to think that it had a lot more to do with pushing 110nm to 550Mhz :)

Linking to fudzilla.com is fun! :smile:

http://www.fudzilla.com/index.php?op...d=345&Itemid=1

oh noes.. I always thought the XT going to do 14k?! AMD is doomed!

Yeah but it's going to score 11400 with 8xAA too :twisted:

And it'll also give you a massage after a hard day and make you a great meal :yes:

rather have a real girl friend do that ;)

"First rule of Fight Club is..." ;)

At this point, leaks of R600 being quite spectacular is the only thing that will stop people throwing up their hands in the air and buying a G80 out of frustration with AMD not providing anything competitive for half a year.

They're getting a 10.6k on the 8800GTX. So the rumors sorta correspond, with the R600 being marginally faster than than the G80, but with vastly great power and cooling requirements, and with a better process and more memory bandwidth. Pretty lame all around I would say.

Useless without AA/AF information.

AMD sometimes looks very paranoid.

Vendors are usually only this strict and paranoid of leaks when there is something either extremely good or extremely bad.....Here's to hoping for extremely good....I'd love to see the GPU industry take yet another big leap forward....

Just about anything can be spectacular from certain points of view...

Now, now, don't go all Obi-wan on us! :lol:

Or are you suggesting niche brilliance somewhere, like say GPGPU, that may or may not be matched with more mainstream brilliance in games?

And resolution, driver settings, Windows version, driver version, and a few other niggly system specs that can potentially impact 3dmark06 scores. There's absolutely nothing to go on there, he might aswell state that G90 gets 20,000 points.

:eek: Maaaaaaaaybe... I'm basing this on the teraflop-in-a-box demo, supposed leaked benchmarks, and Mike Houston's comment that "G80 is more like 330GFlops with user available shader resources" on the gpgpu.org forums. Clearly, R600 is a ton faster in GPGPU apps, but nobody's reporting that kind of difference (>50%) in 3D apps.

Something would have to be holding R600 back if it can't manifest something close this supposed 50% advantage over G80.

The teraflop in a box demo was closed-box. Any potential bottleneck that would be unavoidable in real 3d applications may have been worked around in a toy software loop.

There's all kinds of bottlenecks that could come into play when you consider "the leaked benchmarks" we have so far are mostly either 1) trash or 2) default (no aa/af) 3DM numbers that even if they aren't trash are useless for high-end gpus. GPGPU is quite a different kettle of fish, as R580 vs G7x proved conclusively.

And judging by the specs, I can tell NV30/35 should be at least 50% faster than R9700/9800.

Sigh, I guess fanboys never give up...

To be honest, to compare R600 with NV30 is an insult to NV30, since after all, NV30 is only 4 months later.:lol:

I can't rule out the very likely possibility those benches are garbage.

If it's not that, I don't know how much AA or AF would affect the math throughput in a GPGPU demo like the teraflop one.
Unless R600 is able to utilize those resources as additional FLOPS, the baseline should be 50% higher.

How could a 50% math advantage fail to produce significant performance gains just because AA and AF are off?

Talk about comaring apple to orange: Vector-based flops is not as useful as scalar based flops.

By the way, in theory, NV30 has a 25.6GFLOPS when it comes FP32, and a >50GFLOPS when it comes to FP16, while R300 has a 20.8GFLOPS math power.

I'm comparing theoretical FLOP counts.
If there is a mathematical advantage to R600, some aspect of current apps would have to show it. It has to materialize as a real gain somewhere.

If the reason R600 can't take advantage of its theoretical advantage is that it is suffering from vector utilization issues (we don't know that is true, and some rumors seem to indicate it isn't that), it goes back to my argument that something is holding R600 back, something that might be easier to hide in a closed-box demonstration.

Congratulations. You've just given Fuad his news headline for tomorrow. It'll read something like...

"G90 to hit 20,000 in 3DMark06!

An anonymous source has leaked secret benchmarks that the G90 can hit an AMAZING 20,000 points in 3DMark06! Obviously we can't reveal the source of our information with such an amazing scoop, but we'll keep you updated with additional benchmarks/info as soon as our source reveals them!"


Except with more typos.