View Full Version : Fermi - will gamers get value for money?
FlakMagnet
10-14-09, 09:37 AM
Looking at the recent architecture descriptions of Fermi (and the photo of the die), it looks as if a lot of the chip is taken up by the 512 processing cores (about 2/3 of the surface of the chip).
Are these cores utilised by 'normal' graphics processing or are they only used for GPGPU scenarios (CUDA etc.)?
I'm hoping the cores are also used to run vertex and pixel shaders, and maybe the new tessellation engine, but it sounds as if the 3-stage tessellation is in dedicated hardware.
I'm just thinking that if these cores are only used by CUDA, DirectCompute and OpenCL, then gamers will be paying a lot of money for a lot of hardware they won't necessarily use.
I hope I'm wrong about this.
LydianKnight
10-14-09, 10:11 AM
NVIDIA renamed the cores as 'CUDA cores' but that doesn't mean they're more suitable for GPGPU computation, just they're now optimized to do both, not only gaming, you'll get 2.13x the number of cores inside a GTX 285 (just to name a simple fact), and as the cores are optimized for parallel data processing, if they'll be able to use CUDA more efficiently, I don't see why can't it be used for more fine-grained shader computation, whether it's for vertex or pixel shaders
So will gamers get value for their money? I'd say yes, but it's just my opinion.
FlakMagnet
10-14-09, 01:12 PM
OK, so if the CUDA Cores are also used for vertex and pixel shaders then I'm happy :)
OK, so if the CUDA Cores are also used for vertex and pixel shaders then I'm happy :)
With DX10 they started to use Unified Shaders.
One shader unit, that can run Vertex, Pixel and Geometry shader code.
CaptNKILL
10-14-09, 03:34 PM
Hopefully they don't neuter the fill rate and memory bandwidth to cut costs and focus on GPGPU applications. It could have 50 million shader processors but if it only has 10% more fill rate than a 285 it'll still choke at high resolutions with anti aliasing.
FlakMagnet
10-14-09, 03:51 PM
Agreed. It's going to have to be considerably faster in games than my GTX280 to warrant a purchase. And I want to know that if I do buy one that I'm getting use of all 3 billion transistors.
Does anyone know if the tessellation will be performed by these cores (possibly in CUDA code or C++) or is there dedicated hardware for that? It would make sense to do everything on the cores if possible allowing for fixes/performance improvements to the algorithms if required.
If the card truly is a GPGPU, does that mean that it could potentially support future features (DirectX12 for example) by simply updating the code that it runs? If that is the case, we might end up with a card that is future proof. New DirectX? No problem, just update the drivers and the card can support it natively. People wouldn't 'need' to upgrade because of a newer DirectX - they would upgrade only when they wanted a faster card.
turdhat
10-14-09, 05:01 PM
Agreed. It's going to have to be considerably faster in games than my GTX280 to warrant a purchase. And I want to know that if I do buy one that I'm getting use of all 3 billion transistors.
Does anyone know if the tessellation will be performed by these cores (possibly in CUDA code or C++) or is there dedicated hardware for that? It would make sense to do everything on the cores if possible allowing for fixes/performance improvements to the algorithms if required.
If the card truly is a GPGPU, does that mean that it could potentially support future features (DirectX12 for example) by simply updating the code that it runs? If that is the case, we might end up with a card that is future proof. New DirectX? No problem, just update the drivers and the card can support it natively. People wouldn't 'need' to upgrade because of a newer DirectX - they would upgrade only when they wanted a faster card.
I am running a gtx 280 overclocked to 726core/1439shader/1250mem and every game I throw at it runs smoothly including all the latest titles. Need for speed shift stays around 60fps almost all of the time even at 1920x1080 16xaf and 4xaa. I had a 720p display until a month ago and upgraded my hdtv to 1080p. I though for sure Dirt, Grid and other games would suffer.. Nope, all still smooth. I do have a core2 quad at 3.6 however. Still, I see no reason to upgrade at all. Dont need more than 4axx at 1080p. It looks fine with no jaggies I can see even on a 55 inch screen.
FlakMagnet
10-14-09, 05:35 PM
I've recently finished Batman, and played with everything on max. including the PhysX. Although the game was very playable at 1920x1200, the framerate did drop into the 20's sometimes. So a card that can run at maximum detail with PhysX on the same card would be nice.
I'm also looking forward to playing DiRT 2 in DirectX11 mode as this is promised to be the best looking driving game ever made (a developer quote, but I'm assuming they have seen Forsa 3 and GT5 screenshots). Other DirectX11 games are also going to be launched before the end of the year, and other in the first quarter of next year. I'm particularly looking forward to Alien v Predator which is looking really nice.
I also run quite a few games with 3D enabled. Left 4 Dead looks amazing in true 3D - as does Resident Evil 5. However, enabling 3D really effects the framerate as everything is rendered twice. A faster graphics card would definitely help in that respect.
On my GTX280, Batman with max. everything, PhysX on High and 3D enabled really crawls and is unplayable. I'm sure this is true of other games too. So although the GTX280 is a very fast card, it can't cope in the most extreme situations. I'm hoping the GTX380 will :)
josiahsuarez
10-19-09, 09:57 PM
I also run quite a few games with 3D enabled. Left 4 Dead looks amazing in true 3D - as does Resident Evil 5.
are you referring to the anaglyph red and blue 3d glasses or the glasses with the polarized lenses?
If the card truly is a GPGPU, does that mean that it could potentially support future features (DirectX12 for example) by simply updating the code that it runs? If that is the case, we might end up with a card that is future proof. New DirectX? No problem, just update the drivers and the card can support it natively. People wouldn't 'need' to upgrade because of a newer DirectX - they would upgrade only when they wanted a faster card.
direct X 10.x cards can support many direct X 11 features. they simply are not at the dx 11 standard. this meant that they cant do all the things a dx 11 card can. when a dx 12 card comes out a dx 11 card will not be able to support the full spec because dx 12 will have new feature sets that cant be programed in as some of them are at the hardware level.
at least thats how I understand it. someone correct this if im wrong.
I've recently finished Batman, and played with everything on max. including the PhysX. Although the game was very playable at 1920x1200, the framerate did drop into the 20's sometimes. So a card that can run at maximum detail with PhysX on the same card would be nice.
Some of the Fermis features should come very handy for PhysX and games, I would quess the paraller kernel execution should be perfect for reducing stalls.
That feature alone could improve playability of PhysX titles with a nice margin.
Other features like 2L cache and such can mean improved performance/core for a physx and many other computation heavy task.
If there is no apparent bottlenecks with the design we will have nice 'little' card on our hands. :)
LordJuanlo
10-20-09, 04:49 AM
Or you could use an old, dedicated card for PhysX, like I do and most other people do. Batman clearly states that you need a dedicated card for high detail PhysX
Or you could use an old, dedicated card for PhysX, like I do and most other people do. Batman clearly states that you need a dedicated card for high detail PhysX
True. :D
But I'm sure that you wouldn't mind similar performance with single card or with SLI with all the cards doing some graphics as well. ;)
shadow001
10-26-09, 01:15 PM
Should be interesting in one situation though....Since the GT300 doesn't have a hardware based tesselator within the architecture,and relies on the shader cores to handle that should the game use it,it also means less shaders available for actual shading operations within the game.
The HD5870 packs 2.7 Terraflops of single precision shading power,and a seperate hardware based tesselation unit within the overall architecture,so you still get full shader power even if tesselation is being done in game....Something to think about.
LydianKnight
10-26-09, 02:25 PM
Hmm... that's true... let's just hope the performance drop isn't that much with the use of the load balancing features inside the reworked CUDA cores.
Anyway, it's still bit early to determine a winner given any scenario, until we see some real benchmarks we're left in the dark :zombie:
Should be interesting in one situation though....Since the GT300 doesn't have a hardware based tesselator within the architecture,and relies on the shader cores to handle that should the game use it,it also means less shaders available for actual shading operations within the game.
The HD5870 packs 2.7 Terraflops of single precision shading power,and a seperate hardware based tesselation unit within the overall architecture,so you still get full shader power even if tesselation is being done in game....Something to think about.
If game tesselates polygons small enough I'm pretty sure that atis performance goes down well before it can actually use those ALUs for full potential.
This due to the vertex setup and quad rendering inefficiency with a small polygons.
But yes, it will be interesting to see the performance on both cards. :)
Since the GT300 doesn't have a hardware based tesselator within the architecture
Is that something that has been officially confirmed though?
That they dont have a fixed function tessellator on there?
I´ve seen the talk, but doesn´t that all stem from the abscense of one on slides?
Seems very strange if it is the case.
fixed function HW is much faster then doing it with GS code, which is why noone bothered with Tessellation in DX10, which is entirely possible, but slow and inefficient.
shadow001
10-27-09, 01:20 PM
Is that something that has been officially confirmed though?
That they dont have a fixed function tessellator on there?
I´ve seen the talk, but doesn´t that all stem from the abscense of one on slides?
Seems very strange if it is the case.
fixed function HW is much faster then doing it with GS code, which is why noone bothered with Tessellation in DX10, which is entirely possible, but slow and inefficient.
Anandtech mentioned it on the fermi article published last month,when fermi was officially unveiled.
Anandtech mentioned it on the fermi article published last month,when fermi was officially unveiled.
This one?
http://www.anandtech.com/video/showdoc.aspx?i=3651&p=1
Searched every page for "tes", cant find it.
shadow001
10-27-09, 03:52 PM
This one?
http://www.anandtech.com/video/showdoc.aspx?i=3651&p=1
Searched every page for "tes", cant find it.
My Mistake...It's shown here:
http://www.osnn.net/graphics-cards/98517-fermi-nvidia-next-gen-gcpu-architecture.html
It specifically mentions tesselation support via software.
My Mistake...It's shown here:
http://www.osnn.net/graphics-cards/98517-fermi-nvidia-next-gen-gcpu-architecture.html
It specifically mentions tesselation support via software.
And is that credible then? I mean is that really official?
shadow001
10-27-09, 05:52 PM
And is that credible then? I mean is that really official?
Well the picture in that link of the GT300 shows no tesselation unit either and that comes straight from Nvidia,and here's a partial picture of the ATI cypress GPU.
http://images.anandtech.com/reviews/video/ATI/5870/GE.jpg
Bottom line is that if it had one,and with that particular feature being part of the DX11 standard,i don't think Nvidia would keep it a secret.
InqWoN1776
10-27-09, 06:14 PM
I think the presentation nvidia gave was about the Fermi architecture and how it relates to their Tesla line of products.
josiahsuarez
10-27-09, 11:31 PM
as I understand it tessellation in the Unigine benchmark can fall back to a software implementation if your card doesn't support DX11 (although i think you need to have DX11 installed anyway for the wrapper to function). it will be interesting to see a thorough comparison between DX10 cards, 5870 and Fermi once it's available.
shadow001
10-27-09, 11:53 PM
as I understand it tessellation in the Unigine benchmark can fall back to a software implementation if your card doesn't support DX11 (although i think you need to have DX11 installed anyway for the wrapper to function). it will be interesting to see a thorough comparison between DX10 cards, 5870 and Fermi once it's available.
Well,performance FPS wise,drops ~40 FPS on an HD5870 card at the default benchmark settings,between running in DX10 mode(no tesselation) and DX11 mode,and the HD5870 has a hardware based tesselation unit within the architecture,so it remains to be seen what the GT300 can do there in terms of performance if tesselation is done in software.
That's one of the things i hate about Direct X overall,as it makes supporting the feature set mandatory,but has no bearing on how those features are implemented within the architecture,or has provisions for minimum performance standards for every feature,as there's been plenty of examples in the past of features being nothing more than check box items,and not being designed to have enough performance in the real world.
vBulletin® v3.7.1, Copyright ©2000-2012, Jelsoft Enterprises Ltd.