The GT300/Fermi Thread - Part 2!

[saaya]

well a paper launch means no cards in retail at all...
from what i heard there will be cards for the launch, just not a lot...
then again, there have been a lot of hints that partners actually still dont have those few thousand cards at hand that will supposedly be available at launch... and they will need at least a week to put their own stickers on them, flash their own bios on them, put them in a box with some games and apps and then ship them to distributors and shops... more like 2 weeks... so gigabyte, asus, evga, zotac, msi, bfg etc... they SHOULD all be receiving piles of fermi cards right now...

another weird thing, considering that cebit was 3 weeks before the launch, with supposedly thousands of cards available in retail... its a bit surprising that there were merely 6 cards in total at the show... 6... and at least one of them wasnt even a3 but still a2 silicon... if there will really be a few thousand cards at launch, nvidia and their partners must be rushing things, thats the only thing that could explain this...

[saaya]

Well good at least nVidia is working on parallelizing more of the pipeline.

isnt that what dx10.1/dx11 is about?

Serial would be better, but clocks have been stuck in 600-700 range for half-a-decade

gpus dont exist for that long
r200 and nv20 were both around 250mhz...
ati clocks only increased from 600 to 850mhz in the past years, so its def slowing down... nvidia has decoupled shaders and is clocking them higher, but those clocks havent changed since G80... and who knows how high ati clocks some domains in their gpus...

gpus struggle to break 1ghz just as cpus struggle to break 4ghz...
the future is definately all about parallel work, and both gpus and their code, games, are used to this for over a decade already... so they have a natural advantage in this race...

[Tim]

*snip snip snip

From the same guy PHK on bbs.expreview.com forum.

Don't know what card is, maybe GTX 470 but not sure

That doesn't look too bad if that's the 470. The new mid/high card from nVidia against ATIs high end single GPU. Just a bit late lol.

I think all DX9 and DX10 benches won't matter much. I think DX11 is where it's at, or rather going to be at in the future.

[Dimitriman]

That doesn't look too bad if that's the 470. The new mid/high card from nVidia against ATIs high end single GPU. Just a bit late lol.

I think all DX9 and DX10 benches won't matter much. I think DX11 is where it's at, or rather going to be at in the future.

There is a slight problem with the benchmarks going around demoing GTX470's speed. They are nearly all 2500 resolution which is where the card really pulls very close if not identical scores to 5870. While that is great for the 470, it hides the fact that its still falling behind on lower resolutions. Although overall they should be priced the same IMO since difference is only marginal.

Now.. on the other hand, a 1Ghz clocked 5870 seems to be matching anything Nvdia is throwing out there with Fermi which makes me wonder how AMD will clock and position its refresh 5k series.

PS: If you are thinking that DX11 = Tesselation then maybe there will be differences, but I very much doubt that the results for both cards will differ significantly under DX11. Both are early versions of DX11 cards and by the time its a popular standard they will be replaced by much faster versions.

[informal]

^^ Agree with Dimitriman's post.

[thatdude90210]

There is a slight problem with the benchmarks going around demoing GTX470's speed. They are nearly all 2500 resolution which is where the card really pulls very close if not identical scores to 5870. While that is great for the 470, it hides the fact that its still falling behind on lower resolutions. Although overall they should be priced the same IMO since difference is only marginal.

Resolution of 2560x1600 is Nvidia's best bet. By upping the res and AA, they can make the 1GB 5870 run out of memory, causing stuttering and big hits in framerate.

[eric66]

There is a slight problem with the benchmarks going around demoing GTX470's speed. They are nearly all 2500 resolution which is where the card really pulls very close if not identical scores to 5870. While that is great for the 470, it hides the fact that its still falling behind on lower resolutions. Although overall they should be priced the same IMO since difference is only marginal.

Now.. on the other hand, a 1Ghz clocked 5870 seems to be matching anything Nvdia is throwing out there with Fermi which makes me wonder how AMD will clock and position its refresh 5k series.

PS: If you are thinking that DX11 = Tesselation then maybe there will be differences, but I very much doubt that the results for both cards will differ significantly under DX11. Both are early versions of DX11 cards and by the time its a popular standard they will be replaced by much faster versions.

agreed 470 only closes the gap in 2560 res where 5870 ram isn't enough which doesn't tell true story when 2 gb versions will roll out game will change and i think we will see them even before fermis launch so it will be something like 4870 vs core 192 but gtx 280 being more close to 4870 this time

[MrMojoZ]

i think we will see them even before fermis launch so it will be something like 4870 vs core 192 but gtx 280 being more close to 4870 this time

Which is bad news when your chip costs more to make than the other guys.

[Tim]

There is a slight problem with the benchmarks going around demoing GTX470's speed. They are nearly all 2500 resolution which is where the card really pulls very close if not identical scores to 5870. While that is great for the 470, it hides the fact that its still falling behind on lower resolutions. Although overall they should be priced the same IMO since difference is only marginal.

Now.. on the other hand, a 1Ghz clocked 5870 seems to be matching anything Nvdia is throwing out there with Fermi which makes me wonder how AMD will clock and position its refresh 5k series.

PS: If you are thinking that DX11 = Tesselation then maybe there will be differences, but I very much doubt that the results for both cards will differ significantly under DX11. Both are early versions of DX11 cards and by the time its a popular standard they will be replaced by much faster versions.

I agree with that. It usually takes a generation or two to run a new DX fast. Strange really.

Anyway, the resolution thing is very true.

[***Deimos***]

Crysis2 wont need supercompter

Hello everybody, I'm Mr Obvious:

When Bush started his presidency, there was a crazy mad rush to buy 14"-15" laptops. With the power of Pentium M, the market boomed.

But laptops were very expensive. But, by the time Iraq War was in full swing, prices finally came down and <$1000 student/mobile-business market exploded.

And then some folks tried playing games, like Sims or WOW - it was brutal. So along came 17" latops. But, Fortunately Core2 and HD4xx/GF8/GF9 solved all that. And presto, you got 13" notebook with decent graphics.

But battery life, although improving greatly was still so-so. CULVER fixed that. Now you can get an 11.6" Alienware MX11 with 6000 3DM6 and 4.5hr battery.

The consequences: - virtually all my undergrad friends have notebooks/netbooks.
They are all newer than their desktops. I did an informal survey, and the girls in class prefered notebooks to desktop by HUGE margin.
- current 25+ demographic is still clinging to "PC", but to the new young generation its getting old like rotary phones. Perhaps, the aisles and aisles of $500-$700 notebooks at BestBuy, and PCs hiding somewhere in the back has something to do with this.

Thus, days of GF2 GTS / GF4 are LONG GONE. PC gamer market is still "around".. but certainly not growing like before.
GTX480/ Fermi is niche market product. It will never have broad appeal. Even if production was possible, its very hard to imagine tens of million units sold.

[Tim]

Yet the high end cards continue to boost graphics, and without them, laptops wouldn't have their derivatives. So it doesn't really matter if not tens of millions are sold. It's continuously pushing the industry. Without it, you'd still have Geforce 2 MX onboard in your laptop.

Ps. how offtopic can a post go lol. Plus this is XS, laptop talk is prohibited. j/k

[Chumbucket843]

How easy is to to scale from thousands of threads to millions? Buffers and register files are already enormous.

Problem is that graphics is assumed to be infinitely parallel. But you can only work on about 2 million pixels at a time, before starting next frame. What happens with 20 million triangle tessellation demo running at 800x600... 40 triangles/pixel on avg - which you cant distinguish. Of course we're still some years away.

well, then we will be able to render scenes that are more accurate then our eyes can see. rv870 has ~5MB of registers. memory hierarchy will be a big focus in new architectures. ie. fermi.

But perhaps the greatest challenge is the efficiency ceiling. Some shader applied to block of 15 of 16 pixels is ~94% resource efficient. But extrapolate to dozens of shaders, and 0.94^12 doesn't look so good. GPUs already use an enormous amount of optimizations, from non-sqrt Z calc, to buffer compressions, to normal/bump/stencil maps, mip maps and LOD. It seems only natural that its more and more difficult to improve this fine tuned model.

they have already fixed that problem. pixel shaders run in 32 wide vectors on nvidia and ATi(although ATi can do 5 madds per pixel compared to 1 on nv). you dont have to double the vector width and its actually invisible to the programmer. all they have to do is double the number of SIMDs.

graphics pipelines can always be improved and modified for whatever software demands. thats what the research they do is for.

[LiquidReactor]

There is a slight problem with the benchmarks going around demoing GTX470's speed. They are nearly all 2500 resolution which is where the card really pulls very close if not identical scores to 5870. While that is great for the 470, it hides the fact that its still falling behind on lower resolutions. Although overall they should be priced the same IMO since difference is only marginal.

Now.. on the other hand, a 1Ghz clocked 5870 seems to be matching anything Nvdia is throwing out there with Fermi which makes me wonder how AMD will clock and position its refresh 5k series.

PS: If you are thinking that DX11 = Tesselation then maybe there will be differences, but I very much doubt that the results for both cards will differ significantly under DX11. Both are early versions of DX11 cards and by the time its a popular standard they will be replaced by much faster versions.

Resolution of 2560x1600 is Nvidia's best bet. By upping the res and AA, they can make the 1GB 5870 run out of memory, causing stuttering and big hits in framerate.

ATI is more efficient with memory usage so it shouldnt be as memory capped at that res as you think it would.

[zerazax]

ATI is more efficient with memory usage so it shouldnt be as memory capped at that res as you think it would.

New architecture = all bets are off

[eric66]

ATI is more efficient with memory usage so it shouldnt be as memory capped at that res as you think it would.

what efficiency are you talking about there are graphs from amd showing dramatic increase with 2 gb ram

[Olivon]

what efficiency are you talking about there are graphs from amd showing dramatic increase with 2 gb ram

Slides from AMD are more "crédibles" than nVidia ones ? ... You surprised me ...

[Sly Fox]

what efficiency are you talking about there are graphs from amd showing dramatic increase with 2 gb ram

I think he's referring to a test someone posted here in the 4800 vs GTX 200 series days showing that in the same games with the same settings ATI used less video memory.

AKA more efficient.

Whether that's still the case today or not I don't know.

[RaV[666]]

There is no magical cure for running out of memory.1GB isnt enough for mmst demanding titles/settings in 2560x.
One hint od that is really low minimal fps, that means a card struggled to go through, even if average max aint bad, REALLY low min fps most of the time means low memory or IO bound.
And with a few benchmarks here at 2560x ATI and Nv had similar average fps ,on ati part the min one was below double digits.Low on mem.

To thinking people, another thought, if it wasnt about ATI getting low on memory, they might have used some lower resolutions beside 2560 too :-) .Most gamers ard at 1650x /1920.
Like 95% of leaked scores$20were at 2560x.And the ONE it had 1920x ,fermi wasnt faster...

[***Deimos***]

I remember the "tiny" 256MB on the 36 pixel shader X1950PRO was enough for 1900x1200 4xAA - well for most games anyway.

Then again this was 2006, years before huge memory hoggers like GTA IV.

BTW: To check memory usage why not use rivatuner plugin or something similar rather than assuming 2560x1600 requires 1GB in game X.

[Jowy Atreides]

Let's say, hypothetically, that ATI release a 5870, just reference clocked to 1ghz.

Could nvidia match it?

ie, could they make a higher clocked 480? Or is Fermi totally maxed on clocks and power draw?

[trinibwoy]

<deleted>

[-=DVS=-]

Let's say, hypothetically, that ATI release a 5870, just reference clocked to 1ghz.

Could nvidia match it?

ie, could they make a higher clocked 480? Or is Fermi totally maxed on clocks and power draw?

Its only limitation is probably heat , and we can always increase to 3 slot cooler i would like one .

[Jowy Atreides]

Its only limitation is probably heat , and we can always increase to 3 slot cooler i would like one .

Water cooling
*cool guy face*

[-Sweeper_]

http://translate.google.com.br/trans...n&hl=&ie=UTF-8

hmm 1ghz @ 320-bit sounds nice..

[highoctane]

ie, could they make a higher clocked 480? Or is Fermi totally maxed on clocks and power draw?

Why not, for a std clocked part they have to bin at an avg speed that a majority of gpus will operate for yield & business reasons. There should very realistically be gpu's peppered on the wafer that are capable of higher clocks than the mainstream/std clocks.