Maybe if you didn't quote just part of what I said you would understand ?
Silus said:
Yes, almost 3 years, because only with RV870 did ATI definitely took the performance crown from NVIDIA's "hands".
NVIDIA GF100 & Friends speculation
- Thread starter Arty
- Start date
Maybe if you didn't quote just part of what I said you would understand ?
Yes, almost 3 years, because only with RV870 did ATI definitely took the performance crown from NVIDIA's "hands".
Mintmaster
Veteran
Not at all. 5870 can reach its peak FLOPS rating for mathematically intensive workloads like matrix multiply. If NVidia could figure how to increase its math ability and feed the units as cheaply as ATI, it absolutely would.
Everything is a tradeoff. In RV770, for example, the ALUs occupied around 25% of the die space. They could halve the ALUs, save 12% of the die, and maybe 5% of total board cost, but if those ALUs are the limiting factor in 20% of the total cycles that a GPU is used for gaming, then it loses 10% performance. That's not a good tradeoff. So despite being able to use more than 400 SPs only 20% of the time, it would be a good decision to double it to 800.
To you, ATI's GPUs have more ALUs than it needs. To a smart hardware designer, it's got just the right amount.
What I was trying to tell you is that ATI doesn't have to sacrifice its FLOPS to increase geometry throughput, because it's unrelated to the shader architecture. There's no scaling problem, and they just have to bother doing it. I hope it happens in the next generation for ATI.
It is not FUD. The smaller GF10X chips will show the difference to the smaller RV8X0.
It is FUD, because they are "real tessellation" units. The differences in architecture are irrelevant. Just because one has a fixed tessellation unit, while the other has one per SM, doesn't make it any less real or efficient.
Mintmaster
Veteran
How can you not acknowledge that ATI caught up with RV7xx? It was faster than G92b and crushed GT200 derivatives before they even came out. It clearly overtook NVidia in architecture.
Performance of the top single chip part is nothing more than a business decision. At the very least you should acknowledge the 3870x2 or 4870x2, which were released 4-5 months before their NVidia counterparts.
Last edited by a moderator:
Just wanted to say I'm amused that you can say "It is not FUD" and then immediately follow it up with an assertion that's backed up by nothing. Carry on.
With the 9800GX2 (1 1/2 months behind the 3870x2) nVidia got the performance grown back for nearly 5 months and with the GTX295 for over 10 months.
I'm not going to go the route you went with your first line...
Also, "overtaking in architecture" is irrelevant in regular consumers eyes. It's just like the debate about die sizes. Someone back at [H] even suggested that the die size was very important for the consumer, to which I asked "do you go to a store, wanting to buy a card, and ask about the die size of the chip you are about to buy ?"...it's silly and has no relevance whatsoever...
Now obviously, in consumers that follow architectures and spec sheets, etc, obviously RV770 was better than GT200 and I never said anything that denied that. But the performance crown stayed in NVIDIA's hands during the GT200 vs RV770. As for the X2s, sure, but that's why I said "definitely took the performance crown" in regards to RV870, because with the RV670 X2 and RV770 X2, there was always a counter, with which NVIDIA took it back.
Its not imposible to run tesselation in software. Try out this nvidia dx10 SDK sample caled instanced tesselation. Its pure software tesselation and looks good and runs fine. http://developer.download.nvidia.com/SDK/10.5/direct3d/samples.html#InstancedTessellation
It runs on my radeon too 4850.
Edit : The joke is that this tesselation SDK demo runs better than the ATI tesselation demos. I reach vsynced 60fps even at the max 32 tesselation levels.(thats shitload of geometry)
Last edited by a moderator:
Thanks, but I never said it couldn't be done...
I just said that saying that GF100 doesn't have "real tessellation" is silly at this point.
just a minor rectification of this error - the geforce 3 predated the 8500 by over 6 months. The 8500 was meant to be a geforce 3 killer, but early driver issues meant it took a while to achieve it's true potential. In the end, at very high AF, it was able to challenge the gf4 4200.
edit: meh, didn't notice chalnoth's post, and can't delete
Yeah, the geforce 3 predated the 8500. But the GF3 and the Ti500 are different SKUs. First, the geforce 3 was launched. Than, almost at the same time as the 8500, Nvidia launched the Geforce 3 ti500 (to compete with the 8500) and ti200 (lower cost SKU).
In fact, Nvidia didn´t need the Ti500: in august of that year, a new driver was released wich increased the GF3 performance by 30%.
Just wanted to say I'm amused that you can say "It is not FUD" and then immediately follow it up with an assertion that's backed up by nothing. Carry on.
You can´t deny the fact, that GF100 does not have one fixed tesselation unit.
I must admit that I am really interested to see the smaller and the smallest chips of the Gf100 generation, to see how the tesselation capability scales with the reduction of the SPs, compared to the Radeon chips.
Also with regards to this image :
It is fake , look at the catalyst(or catalytic
I don't know if it's been said already, but this is definitely a fake. Here's our official statement wrt this matter:
http://www.pcgameshardware.de/aid,7...70/480-Benchmarks-im-Umlauf/Grafikkarte/News/
(obviously in german, but google translate will get you the gist).
actually, 10.3a is what we've used. It's internal number is 8.712.3.
You can see our Testing-Lineup (sans Fermi) here:
http://www.pcgameshardware.de/aid,7...orbereitung-auf-GTX-470/480/Grafikkarte/Test/
Last edited by a moderator:
You can´t deny the fact, that GF100 does not have one fixed tesselation unit.
It's easy. It has 15(16) of them.
You can see our Testing-Lineup (sans Fermi) here:
http://www.pcgameshardware.de/aid,7...orbereitung-auf-GTX-470/480/Grafikkarte/Test/
No 5970?
That would make GTX480 look bad, they can't do that
Yes I figured that out, I just want to hear Carsten's pov on it before I say much more.
Hell, it's a graphics board. If I'd want more audio than my onboard solution from realtek, I'd buy an X-Fi
What about open standards support out of the box with certified, regular drivers? I know, i know: It's a software feature. But one I'd really like to have.
Don't get me wrong: I like my HD 5870 and won't exchange it for Fermi anytime soon!
I was under the impressione that especially the last two would need bandwidth, too. And if that's stagnating…
Exaggerated example: how hard is it, to put a couple of thousand DP-ALUs on a die, when you don't have to care about getting them fed with data?
Last edited by a moderator:
Hell, it's a graphics board. If I'd want more audio than my onboard solution from realtek, I'd buy an X-Fi
What about open standards support out of the box with certified, regular drivers? Last time i checked with GPU Caps Viewer from Ozone3D, my HD 5870 didn't have support for Open CL. Actually, that was two minutes ago. But then - I only have Catalyst 10.3a - maybe the WHQL'ed 10.3 fixes that.
Don't get me wrong: I like my HD 5870 and won't exchange it for Fermi anytime soon!
Try installing the STREAM SDK? IIRC it's required for OpenCL atm regardless of driver version
Similar threads
