PDA

View Full Version : NVIDIA GF100 Previews


Pages : 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 [21] 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52

shadow001
03-11-10, 08:39 PM
Heh- why would I watch BluRays on a 23.6" screen? I'll give you better taste in monitors than video cards, but they're kind of wasted with the ATi cards aren't they?:(


When i don't feel like firing up the whole home theater system for it,and can watch them on the PC with a nice set of headphones and still enjoy them at their native resolutions,and even then,still have 2 other screens available to run other applications at the same time,not to mention that i'm also ready for 3D glasses when the standards are set,since the basic requirement is having screens that refresh at 120Hz.


It's not a waste at all,it's prepared for the future for the time being at least.

lee63
03-11-10, 08:50 PM
Here is a little preview. Wonder how much is true. Looks like a 5870 will keep up with a 480 in non tessellation situations.

Not much we haven't seen but at least it all in one thread...nicely done if you ask me.

http://forums.techpowerup.com/showthread.php?t=117000

Rollo
03-11-10, 09:16 PM
When i don't feel like firing up the whole home theater system for it,and can watch them on the PC with a nice set of headphones and still enjoy them at their native resolutions,and even then,still have 2 other screens available to run other applications at the same time,not to mention that i'm also ready for 3D glasses when the standards are set,since the basic requirement is having screens that refresh at 120Hz.


It's not a waste at all,it's prepared for the future for the time being at least.

Maybe ATi's 3d will be ready about the same time as their physics!

http://physxinfo.com/news/2279/amd-and-physx-history-of-the-problem/

shadow001
03-11-10, 09:18 PM
Here is a little preview. Wonder how much is true. Looks like a 5870 will keep up with a 480 in non tessellation situations.

Not much we haven't seen but at least it all in one thread...nicely done if you ask me.

http://forums.techpowerup.com/showthread.php?t=117000


It seems to be a close fight overall,with the winner depending on the game chosen and specific settings played at,which in those cases involves using a 30" LCD display,since it's the only one that natively runs at 2560*1600 resolutions,and even then,running that insane resolution with 8X AA on top,which will stress memory bandwith and the amount of memory onboard the card.


Let's see if some hardware sites actually have the balls to run triple monitor benchmark results,knowing full well that it can't be done with a single Fermi card,while it can be with ATI's cards....It would be histerical if they did that.

shadow001
03-11-10, 09:22 PM
Maybe ATi's 3d will be ready about the same time as their physics!

http://physxinfo.com/news/2279/amd-and-physx-history-of-the-problem/


I'd rather wait for what Samsung,Sony,panasonic's and all other major LCD makers decide on supporting in terms of 3D glasses and if all of them decide to use the same type of shutter glasses and infrared transmitter,then make an informed decision thanks.


Would be histerical if you had to buy another set of glasses to enjoy movies in 3D,if your Nvidia glasses aren't supported,and i'm the type that doesn't go on leaps of faith here....Show me first is my policy.

Razor1
03-11-10, 09:55 PM
Of course,it all depends on how the program uses the resources in the end,but looking at it from the strong points of each architecture,it looks to me like this:


Games using a lot of geometry and memory bandwith = Fermi wins.
Games using a lot of pixel shading and texturing speed = Cypress wins.

Texturing has been a strong suite for pretty much all nV cards in every generation

Pixel shading, so far since the g80, they have done better in more complex shaders, really depends on the shader though. Look at numerous synthetic shader tests.

http://translate.google.com/translate?hl=en&sl=de&u=http://www.computerbase.de/artikel/hardware/grafikkarten/2008/test_radeon_hd_4870_cf_geforce_gtx_260_sli/23/&ei=CLqZS_fBKoP-8Abmk6jKCg&sa=X&oi=translate&ct=result&resnum=1&ved=0CAwQ7gEwAA&prev=/search%3Fq%3Dhd%2B4870%2Bbenchmarks%2Bcomputerbase .de%26hl%3Den

give or take depending on the shader type but as you get to Dx10 shaders, the gtx tends to gain more.

Then you have many other things to consider when you look at the GPU not just shaders and texturing.


So it comes down to which GPU maker made the right call and how developers use those resources in the end,but it's not a knock down victory for 1 GPU,as neither one has the highest performing features for every situation....Both have their respective strong points.


Only because we haven't seen what Fermi can do yet. Can't really compare the HD5xxx to last gen nV, if you want a valid comparison lets use the hd4xxx vs the gt2xx's. Look at above link, its a give or take situation.

And lasts until?....Yup,end of june. April is the date these cards will come out in volume, actually the 470 should have no problem with volume at all.

I think it's more a definitely at this point,with the only question is what kind of refresh will it be,but the least Nvidia should expect is a 10~15% performance improvement over the current HD5870 cards.

That's possible

Johnny C
03-11-10, 09:56 PM
It seems to be a close fight overall,with the winner depending on the game chosen and specific settings played at,which in those cases involves using a 30" LCD display,since it's the only one that natively runs at 2560*1600 resolutions,and even then,running that insane resolution with 8X AA on top,which will stress memory bandwith and the amount of memory onboard the card.


Let's see if some hardware sites actually have the balls to run triple monitor benchmark results,knowing full well that it can't be done with a single Fermi card,while it can be with ATI's cards....It would be histerical if they did that.

That's the point of those benches....framebuffer limited....

shadow001
03-11-10, 10:37 PM
Texturing has been a strong suite for pretty much all nV cards in every generation

Pixel shading, so far since the g80, they have done better in more complex shaders, really depends on the shader though. Look at numerous synthetic shader tests.

http://translate.google.com/translate?hl=en&sl=de&u=http://www.computerbase.de/artikel/hardware/grafikkarten/2008/test_radeon_hd_4870_cf_geforce_gtx_260_sli/23/&ei=CLqZS_fBKoP-8Abmk6jKCg&sa=X&oi=translate&ct=result&resnum=1&ved=0CAwQ7gEwAA&prev=/search%3Fq%3Dhd%2B4870%2Bbenchmarks%2Bcomputerbase .de%26hl%3Den

give or take depending on the shader type but as you get to Dx10 shaders, the gtx tends to gain more.

Then you have many other things to consider when you look at the GPU not just shaders and texturing.


Of course not,but this time around,and for once,it seems that ATI's chip has more texturing power,as Cypress has 80 texture units versus Fermi,wich has 64 texture units,and Cypress runs it's texture units at the same speed as the rest of the core(850mhz),while Fermi will run theirs at whatever speed Fermi runs at(650Mhz???),as only the shaders are double pumped,not the rest of the core.


The same goes for shading power,with Cypress having 2.7 terraflops worth of single precision math at 850Mhz,while Fermi will be no where near that,even for single precision math,as it's rumored they were aiming for 750 Gigaflops for dual precision math,and the tech documents provided by Nvidia when they first presented fermi,states that dual precision runs at half the performance of single precision,so that comes to 1.5 terraflops maximum theoretical for single precision math.


In short,just like Fermi is extremely overpowered for geometry in gaming scenarios,Cypress goes the other way around and really goes nuts on raw shading power,at least in theoretical maximums,and my point is,neither one has the edge in all aspects that define overall gaming performance,as they can only be fast enough relative to the slowest part within either GPU architecture,and all depending on the workload presented by the software itself.


It's becoming a very complex issue to evaluate either architecture,and running the currently available software only provides a partial understanding of what either one is really able to do in practical terms,but it's unfortunately that same software,which doesn't really take advantage of either one all that well,that we'll use the define the winner and the loser....Not fair,but that's life.

XMAN52373
03-11-10, 10:54 PM
Of course not,but this time around,and for once,it seems that ATI's chip has more texturing power,as Cypress has 80 texture units versus Fermi,wich has 64 texture units,and Cypress runs it's texture units at the same speed as the rest of the core(850mhz),while Fermi will run theirs at whatever speed Fermi runs at(650Mhz???),as only the shaders are double pumped,not the rest of the core.


The same goes for shading power,with Cypress having 2.7 terraflops worth of single precision math at 850Mhz,while Fermi will be no where near that,even for single precision math,as it's rumored they were aiming for 750 Gigaflops for dual precision math,and the tech documents provided by Nvidia when they first presented fermi,states that dual precision runs at half the performance of single precision,so that comes to 1.5 terraflops maximum theoretical for single precision math.


In short,just like Fermi is extremely overpowered for geometry in gaming scenarios,Cypress goes the other way around and really goes nuts on raw shading power,at least in theoretical maximums,and my point is,neither one has the edge in all aspects that define overall gaming performance,as they can only be fast enough relative to the slowest part within either GPU architecture,and all depending on the workload presented by the software itself.


It's becoming a very complex issue to evaluate either architecture,and running the currently available software only provides a partial understanding of what either one is really able to do in practical terms,but it's unfortunately that same software,which doesn't really take advantage of either one all that well,that we'll use the define the winner and the loser....Not fair,but that's life.

Every generation ATI has had the lead over Nvidia in SP performance. Alot of good it has done them. You can't use that as a metric.

Razor1
03-11-10, 11:43 PM
Of course not,but this time around,and for once,it seems that ATI's chip has more texturing power,as Cypress has 80 texture units versus Fermi,wich has 64 texture units,and Cypress runs it's texture units at the same speed as the rest of the core(850mhz),while Fermi will run theirs at whatever speed Fermi runs at(650Mhz???),as only the shaders are double pumped,not the rest of the core.


You seem to forget about efficiency, Fermi's new TMU's are more efficient, with x16 AF the g80 doesn't take much of hit, so do we need more texturing power then that?

The same goes for shading power,with Cypress having 2.7 terraflops worth of single precision math at 850Mhz,while Fermi will be no where near that,even for single precision math,as it's rumored they were aiming for 750 Gigaflops for dual precision math,and the tech documents provided by Nvidia when they first presented fermi,states that dual precision runs at half the performance of single precision,so that comes to 1.5 terraflops maximum theoretical for single precision math.

Flops don't mean much if those flops are used in an efficient manner, the g80 had half the flops of the r600 or the rv670, the g200 had half the flops of the rv770, again efficiency is more important. (keep in mind the missing mul, which wasn't used at all)


In short,just like Fermi is extremely overpowered for geometry in gaming scenarios,Cypress goes the other way around and really goes nuts on raw shading power,at least in theoretical maximums,and my point is,neither one has the edge in all aspects that define overall gaming performance,as they can only be fast enough relative to the slowest part within either GPU architecture,and all depending on the workload presented by the software itself.


You don't seem to understand the architecture differences, I suggest you read some of the white papers and B3D articles, they go in depth into this. Also I suggest you look at comparative analysis of the two architectures as well before you draw conclusions based on raw numbers.


It's becoming a very complex issue to evaluate either architecture,and running the currently available software only provides a partial understanding of what either one is really able to do in practical terms,but it's unfortunately that same software,which doesn't really take advantage of either one all that well,that we'll use the define the winner and the loser....Not fair,but that's life.


Its not very hard to evaluate the architecture if you look at how different synthetic tests effect the gpu's. That is why synthetics are good, they stress specific parts of the gpu's that can show you what part of the gpu is strong or weak. How do you think developers optimize different paths? When testing engines we look for certain bottlenecks in the engine (profiling) and then go from there, samethings can be seen with synthetic tests, they stress the gpu's in a specific way where by you know exactly what you are testing for.

lee63
03-12-10, 12:16 AM
Palit :D

http://www.tomshardware.com/news/palit-nvidia-geforce-fermi-gf100,9860.html

shadow001
03-12-10, 12:25 AM
Every generation ATI has had the lead over Nvidia in SP performance. Alot of good it has done them. You can't use that as a metric.


And this time,in texturing too,which would be the first time it's happened in a long time...The original R600 came up way short in that dept with only 16 texture units,and the RV770 took that all the way to 40 units and Cypress doubled that again to 80 units.


They actually doubled shader power,texturing and raw fillrate with Cypress in one shot,and to top it off,kept the same core clock speeds relative to the RV790 chips(850 Mhz),all the while adding about 1.3 billion transistors on top and added DX11 support with Cypress....I'd call that a large improvement in every aspect,no matter how you cut it.

shadow001
03-12-10, 12:45 AM
You seem to forget about efficiency, Fermi's new TMU's are more efficient, with x16 AF the g80 doesn't take much of hit, so do we need more texturing power then that?


It's never been a question if we need more than that,but rather does your competitor offer even more still,and we live in a society where more is always better,and we still have to see how 16AF performance between Fermi and Cypress turns out.



Flops don't mean much if those flops are used in an efficient manner, the g80 had half the flops of the r600 or the rv670, the g200 had half the flops of the rv770, again efficiency is more important. (keep in mind the missing mul, which wasn't used at all)


It all comes down to the software actually using all that shader power,which would benefit ATI's architecture of course,just like Nvidia is going to push hard on the tesselation front to convince users it's the card to buy,even if games never really use all the tesselation ability that Fermi can pull off.


It's the same situation for both ATI and Nvidia,only for different strong points in each GPU(Shader power on cypress and tesselation power on fermi).




Its not very hard to evaluate the architecture if you look at how different synthetic tests effect the gpu's. That is why synthetics are good, they stress specific parts of the gpu's that can show you what part of the gpu is strong or weak. How do you think developers optimize different paths? When testing engines we look for certain bottlenecks in the engine (profiling) and then go from there, samethings can be seen with synthetic tests, they stress the gpu's in a specific way where by you know exactly what you are testing for.


No synthetic test pushes either card hard enough anyhow,as even the heaven demo,when running on my system,in DX11 mode,can average over 100 Fps with tesselation enabled at 1920*1200 and still over 80 FPS with 4X AA and 16X AF on top.


And in another synthetic test(3D mark vantage)it still shows the HD5870 beating the GTX470 at the extreme settings(1920*1200 with 4X AA and full post processing effects).


So which synthetic test are you refering to exactly here?....What we need is a new generation of benchmarking software built to exploit both DX11 and really push even the latest setups to the breaking point performance wise,even with Multi GPU setups,with no backups to DX9 at all,as the heaven demo still has the DX9 path in it.

shadow001
03-12-10, 01:00 AM
Here's a link to the tech report review on fermi,with the specifications on it if it manages to reach a 725 Mhz base clock,and look at the raw fillrates and texturing rates between both fermi and Cypress:


http://www.techreport.com/articles.x/18332/5


Scroll down to see the chart and even the overall article is quite informative,and it shows that texturing rate for fermi didn't really improve compared to the GT200 GPU,and that's if Fermi hits 725mhz on the core.

Enrico_be
03-12-10, 01:56 AM
- http://gathering.tweakers.net/forum/list_message/33626051#33626051 :)

Vardant
03-12-10, 02:57 AM
for launch date are prepared 60 000 cards, in next weeks will be more!
There goes the 5k figure ;)

Iruwen
03-12-10, 04:21 AM
The thing about AMD's model for this and the previous generation is they are looking at scalable solutions that can be increased/decreased in complexity/performance, relatively easily. Nvidia's previous and upcoming model does not allow this flexibility, but they still typically deliver top to bottom market coverage.

This is wrong. Fermi is the most scalable GPU architecture that has ever been built. This was one of the things Nvidia heavily focused on during development. And they do since the NV40. Fermi contains several rasterizers instead of one large unit for example. Scalability simply is a requirement when moving from specialized graphics rendering hardware to multi purpose computing systems.

Iruwen
03-12-10, 04:53 AM
http://www.techreport.com/articles.x/18332/5

Scroll down to see the chart and even the overall article is quite informative,and it shows that texturing rate for fermi didn't really improve compared to the GT200 GPU,and that's if Fermi hits 725mhz on the core.

Eventually texturing simply isn't a bottleneck in real world situations, so there was no need for improvements.

Rollo
03-12-10, 06:36 AM
I'd rather wait for what Samsung,Sony,panasonic's and all other major LCD makers decide on supporting in terms of 3D glasses and if all of them decide to use the same type of shutter glasses and infrared transmitter,then make an informed decision thanks.


Would be histerical if you had to buy another set of glasses to enjoy movies in 3D,if your Nvidia glasses aren't supported,and i'm the type that doesn't go on leaps of faith here....Show me first is my policy.

Obviously you ARE the type that goes on leaps of faith, you bought 3d monitors for video cards that not only have no 3d support, but their manufacturer won't be providing it as well. They just hope some 3rd party vendors will help them out, because they're too broke or inept to develop the tech themselves.

Sort of like with GPU physics- they assumed Havok or Ageia would help them out because they make video cards, and look how that turned out.

I don't doubt there will be 3d movie viewing with ATi, but I think the gaming support will lag. NVIDIA has the dev relations to push innovation and compatibility in the gaming market.

What 3rd party shutter glasses vendor can say the same? <crickets>

For example, at GDC 2010 it was just announced UE3 engines games will now be optomized for 3d Vision.

So let's see, NVIDIA has AA, PhysX, 3d Vision support on UE3- what does ATi have? <crickets>

Buying ATi parts in general is a "leap of faith"- the buyer is left hoping someone will work with ATi someday and it never seems to happen. Then the buyer is left to post "I support open standards!" and play the same way we were 10 years ago.

Iruwen
03-12-10, 07:19 AM
Buying ATi parts in general is a "leap of faith"- the buyer is left hoping someone will work with ATi someday and it never seems to happen.

Like Resident Evil 5 (nana2)

scnr

XMAN52373
03-12-10, 08:26 AM
And this time,in texturing too,which would be the first time it's happened in a long time...The original R600 came up way short in that dept with only 16 texture units,and the RV770 took that all the way to 40 units and Cypress doubled that again to 80 units.


They actually doubled shader power,texturing and raw fillrate with Cypress in one shot,and to top it off,kept the same core clock speeds relative to the RV790 chips(850 Mhz),all the while adding about 1.3 billion transistors on top and added DX11 support with Cypress....I'd call that a large improvement in every aspect,no matter how you cut it.

And they essentials did the same thing R600/670 > R770 and it didn't give them any real performance advantage. It sure looked good in the synthetic benches, but did **** in real world compared to Nvidia. Here we are again with ATI doing the same as before and as before chip to chip(ie single gpu vs single gpu) it isn't working to their advantage again in actual gaming performance. Some thing to concider concerning Fermi. The 8 or 12 ROPs(depending on the card) over 32, aren't really being used for texturing but rather for AA performance. If AA isn't being used, they essentially sit idle. As Razor1 suggested, go read up on the white papers for Fermi.

There is also the fat each generation ATI has had 3x the shaders, but never had a shader based performance lead except for the synthetic based benches. efficentcy is a bitch aint it?

Razor1
03-12-10, 08:51 AM
Here's a link to the tech report review on fermi,with the specifications on it if it manages to reach a 725 Mhz base clock,and look at the raw fillrates and texturing rates between both fermi and Cypress:


http://www.techreport.com/articles.x/18332/5


Scroll down to see the chart and even the overall article is quite informative,and it shows that texturing rate for fermi didn't really improve compared to the GT200 GPU,and that's if Fermi hits 725mhz on the core.

yes it doesn't improve on the raw numbers but its just like ATi's 8x AA performance edge last gen, although the gt200's had higher bandwidth, they weren't able to use as efficient as ATi's counterparts. For the most part ever single part of Fermi has been rebuilt, this isn't a simple extrapolate numbers for last gen situation, until we start seeing some synthetic tests, or real world benches to assume performance based on the chip stats, is pretty much pointless.

Iruwen
03-12-10, 08:57 AM
"Efficiency" meant worse quality in this case though.

Razor1
03-12-10, 10:01 AM
It's never been a question if we need more than that,but rather does your competitor offer even more still,and we live in a society where more is always better,and we still have to see how 16AF performance between Fermi and Cypress turns out.


That really doesn't matter much, most guys that know about the numbers, know to research reviews to make their purchases, most people just care about end performance anyhow. If it was a numbers game, AMD should be selling head over heals since the r600 with the sheer number of shader units on their GPU's.

It all comes down to the software actually using all that shader power,which would benefit ATI's architecture of course,just like Nvidia is going to push hard on the tesselation front to convince users it's the card to buy,even if games never really use all the tesselation ability that Fermi can pull off.


Tessellation is only part of Fermi's performance advantage, the others are still unknown, ATi has pushed hard on tessellation too, which seems to be a mistake now since Fermi will have an advantage here.


It's the same situation for both ATI and Nvidia,only for different strong points in each GPU(Shader power on cypress and tesselation power on fermi).

So far what I have seen nV's focus on their GPU's tend to be more balanced and focused on today's games. Lets have a look at the r520, it is a good chip but it was unbalanced with great AA performance but lacked shader performance against the 7800, then they went to the r580, huge shader performance but we didn't see any games that needed that till next gen cards came out, and that performance was not really seen either. This is where nV tends to have a more consumer centric philosophy when designing their GPU's, ATi at the time tried to market their cards as more advanced and capable of doing more, and they were the truth, but at the end reviews weren't able to show the extent of how much better they were in future applications, since the games out at the time couldn't take advantage of the r580 shader core.

No synthetic test pushes either card hard enough anyhow,as even the heaven demo,when running on my system,in DX11 mode,can average over 100 Fps with tesselation enabled at 1920*1200 and still over 80 FPS with 4X AA and 16X AF on top.

Synthetics push the cards on certain portions, that's the point of synthetics. I'm not talking about game engines like unigine or 3dmark, I'm talking about individual tests, that stress the shader units depending on shader length, texture look ups, vertex texture fetches things like that.

And in another synthetic test(3D mark vantage)it still shows the HD5870 beating the GTX470 at the extreme settings(1920*1200 with 4X AA and full post processing effects).


Once you add AA in there, now you are polluting the test, you are now stressing 2 parts of the card (well actually more since you doing alot more Vantage, its acutally a game engine simulation), how are you going to determine what part is being bottlenecked without a starting point? You have a identify what you are looking for and then proceed from that point. You can't just take this benchmark and this is what it is doing without a reference point. More interested in individual shader tests, fillrate tests, etc. Use those as a starting point then go from there.


So which synthetic test are you refering to exactly here?....What we need is a new generation of benchmarking software built to exploit both DX11 and really push even the latest setups to the breaking point performance wise,even with Multi GPU setups,with no backups to DX9 at all,as the heaven demo still has the DX9 path in it.

You have to break the performance to see where the bottlenecks lie. There are many synthetic tests out there, ixbtlabs and computerbase.de are review sites that uses their own synthetic tests, but there are others too that use like shadermark.

shadow001
03-12-10, 11:42 AM
Obviously you ARE the type that goes on leaps of faith, you bought 3d monitors for video cards that not only have no 3d support, but their manufacturer won't be providing it as well. They just hope some 3rd party vendors will help them out, because they're too broke or inept to develop the tech themselves.

Sort of like with GPU physics- they assumed Havok or Ageia would help them out because they make video cards, and look how that turned out.

I don't doubt there will be 3d movie viewing with ATi, but I think the gaming support will lag. NVIDIA has the dev relations to push innovation and compatibility in the gaming market.

What 3rd party shutter glasses vendor can say the same? <crickets>

For example, at GDC 2010 it was just announced UE3 engines games will now be optomized for 3d Vision.

So let's see, NVIDIA has AA, PhysX, 3d Vision support on UE3- what does ATi have? <crickets>

Buying ATi parts in general is a "leap of faith"- the buyer is left hoping someone will work with ATi someday and it never seems to happen. Then the buyer is left to post "I support open standards!" and play the same way we were 10 years ago.


Maybe it's a question of seeing what's on the market,and then seeing what there is to support from ATI....First instance,it seems that sony system has technology within the glasses themselves that can adapt for the seating position and distance relative to the screen itself.


Do Nvidia's current glasses have that?


And playing on 3 screens,already supporting out of the box at least 30 games,with no need for patches from developers and doing that while only needing a single card,is playing games like they were 10 years ago right??? :p


That's already more than 3X the amount of games,compared to the amount there is for GPU accelerated physics and the cards have only been released a few months.....GPU physics has been available for 2 years now from Nvidia.