skeletor
Chosen of the Omnissiah
HD 7950 > GTX 660 Ti.
Finally someone who tried TXAA out. +1.Please don't make TXAA so much of magic word since saying "TXAA offers much better AA" is very subjective and depends on person to person. There are people out there who prefer MSAA and dont care about TXAA since it actually makes graphics plastici instead of realistic. The amount of blur TXAA gives is actually equal to that used in film CGI. The problem with that is that that blur is part of the "experience" and the image quality of film running at 24 FPS. So it is really bad to see it on a game that is NOT film, nor is it running at 24 FPS
Nica One.. 
not everyone is facing those problems its only them..A driver update to reduce Radeon frame times - The Tech Report - Page 1
There is something wrong with sapphire cards and this beta driver
maybe they were sent faulty card
S.S gadgets
Broken In
But why AMD Cards are slow at video processing??
Actually most of the guys depend on the features of Nvidia cards like phyx effects,CUDA and all stuffs...
skeletor
Chosen of the Omnissiah
1) PhysX is dead.But why AMD Cards are slow at video processing??
Actually most of the guys depend on the features of Nvidia cards like phyx effects,CUDA and all stuffs...
2) AMD Cards are way way faster at Compute. nVidia is very slow at it.
3) nVidia cards are faster at video transcoding because they include a better fixed function video transcoder. But using it (and even Quick Sync) doesn't give you control. Most people use Handbrake on their CPU.
4) nVidia cards are painfully slow at handling 3ds Max viewports. AMD is again much faster at it.
Things mentioned in this review are just select cases except Luxmark which gives the right picture overall i.e. Kepler being slower than Fermi.:
OpenCL: GPGPU Benchmarks : GeForce GTX 660 Ti Review: Nvidia's Trickle-Down Keplernomics
*www.tomshardware.com/reviews/geforce-gtx-660-ti-benchmark-review,3279-13.html
OpenCL: Image Processing (Basemark CL) : GeForce GTX 660 Ti Review: Nvidia's Trickle-Down Keplernomics
OpenCL: Video Processing (Basemark CL) : GeForce GTX 660 Ti Review: Nvidia's Trickle-Down Keplernomics
S.S gadgets
Broken In
Thanks ico
Thanks ico
Thanks ico
vickybat
I am the night...I am...
Newer games supporting physx-
1. Metro last light(upcoming)
2. Hawken (free to play)
3. Lost planet 3(upcoming)
4. BorderlandsII
5. Arma 3(upcoming)
Most new games that will run on the upcoming unreal engine 4 has a high chance of using PhysX including the next batman game.
So its not as dead as people think. Its very much alive and well.
Calling it a deal maker or deal breaker is a different thing, but calling it obsolete is far from truth.
Talk about GPGPU, applications using CUDA are targeted towards supercomputers. Most advanced supercomputers in the world including "Titan" at oakridge,and "CRAY" all use CUDA based applications for monitoring weather, protein folding and a lot of other stuff. CUDA libraries are extremely vast and is an overkill for general purpose use.
Open-cl finds itself in consumer apps as its easier to develop. Nvidia supports both of them but its kepler hardware lacks the fundamental units for vector processing. GCN architecture thus performs great in open-cl apps. However the ones used in the above supercomputers (k20x) are built on GK110 which is nvidia's newest iteration that supports compute. There are reports of them to be faster in compute than GCN based workstation cards.
The next kepler family of gpu's will support compute.
Before debating about compute performance, the user has to ask himself/herself whether they are actually going to use apps that harness gpu's compute ability
in open-cl codepath. In games, it does not matter at all coz both gpu's have enough resources to get directcompute based jobs done.
1. Metro last light(upcoming)
2. Hawken (free to play)
3. Lost planet 3(upcoming)
4. BorderlandsII
5. Arma 3(upcoming)
Most new games that will run on the upcoming unreal engine 4 has a high chance of using PhysX including the next batman game.
So its not as dead as people think. Its very much alive and well.
Calling it a deal maker or deal breaker is a different thing, but calling it obsolete is far from truth.
Talk about GPGPU, applications using CUDA are targeted towards supercomputers. Most advanced supercomputers in the world including "Titan" at oakridge,and "CRAY" all use CUDA based applications for monitoring weather, protein folding and a lot of other stuff. CUDA libraries are extremely vast and is an overkill for general purpose use.
Open-cl finds itself in consumer apps as its easier to develop. Nvidia supports both of them but its kepler hardware lacks the fundamental units for vector processing. GCN architecture thus performs great in open-cl apps. However the ones used in the above supercomputers (k20x) are built on GK110 which is nvidia's newest iteration that supports compute. There are reports of them to be faster in compute than GCN based workstation cards.
The next kepler family of gpu's will support compute.
Before debating about compute performance, the user has to ask himself/herself whether they are actually going to use apps that harness gpu's compute ability
in open-cl codepath. In games, it does not matter at all coz both gpu's have enough resources to get directcompute based jobs done.
Last edited:
techdabangg
Broken In
Hmmm.... Nvidia fanbouys some
1. We are talking about current gen GPUs and not next gen. So "The next Kepler family of gpu's will support compute" isn't valid argument here. Apart from this its not like Kepler does not support Direct Compute, its just that nVidia forgot to optimize scheduler for compute performance, which they'll do for upcoming Kepler Refresh.
2. Agreed about PhysX being not dead, but the real question is - is it really necessary? I mean, I've played batman series on HD7970GHz (without PhysX) and GTX680 (with PhysX). The visual oomphs which PhysX gives isn't that much great. In fact, in Batman series its just some smoke in some corridors or new-papers cuttings going helter-scalter. Its just a marketing ploy by nvidia for PCs. Compare Batman AC between XBOX360, which uses ATI xenos (no-way it supports PhysX libraries) and PC (with nvidia 660Ti). You'll see that the smoke/papers etc, physX effects are present in Xbox version without PhysX support. For PC they've just been omitted for non-nvidia cards. Why and How? Its a developer-nvidia nexus.
Nvidia has had upper hand in marketing tricks over AMD/ATI always.
3. Direct Compute is being used in many games as of now as well, so it does matter in games as well. e.g. Civillization V -
*images.anandtech.com/graphs/graph6025/47486.png
@vickybat - Please don't take me in any wrong way. I'm just putting in here what I've experienced myself. I do test products extensively for a 3rd party multi-national technology firm.
1. We are talking about current gen GPUs and not next gen. So "The next Kepler family of gpu's will support compute" isn't valid argument here. Apart from this its not like Kepler does not support Direct Compute, its just that nVidia forgot to optimize scheduler for compute performance, which they'll do for upcoming Kepler Refresh.
2. Agreed about PhysX being not dead, but the real question is - is it really necessary? I mean, I've played batman series on HD7970GHz (without PhysX) and GTX680 (with PhysX). The visual oomphs which PhysX gives isn't that much great. In fact, in Batman series its just some smoke in some corridors or new-papers cuttings going helter-scalter. Its just a marketing ploy by nvidia for PCs. Compare Batman AC between XBOX360, which uses ATI xenos (no-way it supports PhysX libraries) and PC (with nvidia 660Ti). You'll see that the smoke/papers etc, physX effects are present in Xbox version without PhysX support. For PC they've just been omitted for non-nvidia cards. Why and How? Its a developer-nvidia nexus.
Nvidia has had upper hand in marketing tricks over AMD/ATI always.
3. Direct Compute is being used in many games as of now as well, so it does matter in games as well. e.g. Civillization V -
*images.anandtech.com/graphs/graph6025/47486.png
@vickybat - Please don't take me in any wrong way. I'm just putting in here what I've experienced myself. I do test products extensively for a 3rd party multi-national technology firm.
Last edited:
Cilus
laborare est orare
1st of all, Please don't bring some Super Computer Design here to prove Nvidia's supremacy. Is OP going to buy a Tesla based GPU? No. So point of discussing how they perform. Assignment modules are missing in Kepler gaming card because they generate lots of Heat for gaming card which needs huge amount of real time processing, not because this time Nvidia wanted AMD to win by providing inferior chip.
And GCN performs well in any kind of compute performance than any Gaming card from Nvidia. So keeping super computer apart....it is a very good buy.
Regarding CUDA and OpenCL, Vickybat, you are getting the wrong picture. It is not like CUDA is only for super computer and OPENCL is only for general use. CUDA is not a programming language, it is the name of the GPU stream processor architecture of Nvidia and OPenCL language is also used in Super Compute based applications, it is not like that CUDA is the only GPU coding methodology. And just bringing Titan doesn't prove that Nvidia is superior in Super Computing. Super Comute performance mainly depends upon the number of devices connected in chain, not the architecture of each of the components. Currently Titan has the highest number of Workstation GPU attached with it and that's why it is fastest, not because they use only Nvidia hardware.
And CUDA is just for Super computing and too vast for normal usage.... from where did you get the idea? CUDA, OpenCL and DirectCompute has more than 60% common library functions but handles hardware differently. CUDA has restriction on running only on Nvidia Unified Stream Architecture, OpenCL can target any SIMD (Single Instruction Multiple Data) based architecture, irrespective of their nature and vendor. It can even target an APU or a Multi-Core Processor.
OPENCL is not at all any consumer app development API, it is more useful than CUDA even for designing super computer apps. Don't write anything what you just think without getting the fact correct. Buddy, OpenCL offers far better portability than CUDA and if coded properly it performs as good as CUDA + the added benefit of running it on any SIMD design. Just giving the example of Titan doesn't prove than Nvidia is better in super computing...they are in that business for long time...that's all.
And GCN performs well in any kind of compute performance than any Gaming card from Nvidia. So keeping super computer apart....it is a very good buy.
Regarding CUDA and OpenCL, Vickybat, you are getting the wrong picture. It is not like CUDA is only for super computer and OPENCL is only for general use. CUDA is not a programming language, it is the name of the GPU stream processor architecture of Nvidia and OPenCL language is also used in Super Compute based applications, it is not like that CUDA is the only GPU coding methodology. And just bringing Titan doesn't prove that Nvidia is superior in Super Computing. Super Comute performance mainly depends upon the number of devices connected in chain, not the architecture of each of the components. Currently Titan has the highest number of Workstation GPU attached with it and that's why it is fastest, not because they use only Nvidia hardware.
And CUDA is just for Super computing and too vast for normal usage.... from where did you get the idea? CUDA, OpenCL and DirectCompute has more than 60% common library functions but handles hardware differently. CUDA has restriction on running only on Nvidia Unified Stream Architecture, OpenCL can target any SIMD (Single Instruction Multiple Data) based architecture, irrespective of their nature and vendor. It can even target an APU or a Multi-Core Processor.
OPENCL is not at all any consumer app development API, it is more useful than CUDA even for designing super computer apps. Don't write anything what you just think without getting the fact correct. Buddy, OpenCL offers far better portability than CUDA and if coded properly it performs as good as CUDA + the added benefit of running it on any SIMD design. Just giving the example of Titan doesn't prove than Nvidia is better in super computing...they are in that business for long time...that's all.
vickybat
I am the night...I am...
Hmmm.... Nvidia fanbouys some
Nope no fanboy's here.
Well we are talking about current GPU's here. GK110 is already out in the form of K20X. Its just limited for a specific usage set now and commercial cards are a while away from getting launched.
Talk about optimizing scheduler, you are partially correct there.
HPC or high performance computing is the key what we term as compute performance or GPGPU computing. The logical units involved here should be capable of working on double precision arithmetic (integer or float). GPU's highly parallel architecture makes it the best for incorporating multiple units that compute on double precision arithmetic. GCN also does this and so does intel's knight's corner based co-processors.
Now kepler gk104 (gtx 680 and its derivatives), did not have these units that helps in computation of DP arithmetic and they were handled by conventional streaming units (cuda cores) while instructions scheduled by warp schedulers. Each SMX, has 4 warp schedulers. Warp is nothing but a group of 32 instructions to be scheduled for dispatch and execution. So an SMX can dispatch 32x4 warps to all its execution units.
Now the key in GK110, is that the double precision instructions are grouped together with single precision instructions in common and dispatched so that the cuda cores can work on Single ones whereas the DP units can work on double precision instructions only. Thus its compute performance is seen to be much much higher as it has physical units that assists compute. GK104 did not have these.
Commercial GK110 has 64 DP units in a single SMX and has a total of 15 SMX. The general version might have lesser no. of DP units per smx, lets say 32. Kepler GK104 has zero DP units.
So its not only the scheduler but also due to actual physical execution units that assist compute.
Refer the below diagram coz it has both GK110 & GK 104:
*i.imgur.com/JmzcEEz.png
*i.imgur.com/CLmmCan.png
*i.imgur.com/CLmmCan.png
I would say, yes. Physics is necessary and Physx is simply a GPU implementation. The engine is similar to other cpu based physics engine like havok and bullet. In fact the library set of Physx is quite similar with havoc but run on gpu instead. Talk about visual oomphs, you don't expect physics code to do ray tracing ambient occlusion or HDR do you? Its only meant to be in the game engine for the behavior of objects in the game world complying with laws of physics in the real world as close as possible. The flying of newspaper, scattering of ash particles, bullets ricocheting, explosions, movement of cloth etc are the main aspects that a physics engine takes care of. All these things definitely add up to the visual oomph.
Besides batman arkham city is by far the best games to implement Physx and looks way way better than xbox 360. Well the physics engine used in consoles is different. They have engines which can be used by them and thus both ps3 and xbox 360 versions show physics effects. But the pc version employs nvidia's physx engine which has been locked up from amd cards .
But arkham city's pc graphics and physics effects are a lot better than its console counterparts. The first scene itself will tell you. Console version has a much toned down version of ash scattering effects throughout the city along with hail. I've compared in game console videos with my friend's 6870 cf + GT240 based system. Arkham city just looks amazing in pc with physx effects. Its not like amd cannot do these, but it has been deliberately shut for them and we know the reason.
One thing i'll agree on is that current physx implementation in games doesn't propel the necessity of physics code running in a gpu as till date, there has been no effects which can't be done by other physics engines run on cpu. So its still a work in progress and maybe in future we might see multiple physics engines developed to run on a gpu with huge library set that can harness gpu's high parallel computing abilities.
I agree that direct compute is used in many games. What i meant is that in game implementation of direct compute do not harness a gpu's absolute compute potential.
Basically directcompute is used for implementation ambient occlusion in games by ray tracing methods. Farcry 3, hitman absolution , sleeping dogs are some great titles that use directcompute.
But we haven't seen GCN cards to have edge on kepler in AO performance in these games. Kepler performs toe to toe with GCN and can apply the highest AO (HDAO) while delivering similar frames.
The cuda cores of kepler have enough horse power to utilize direct compute and it isn't getting crippled at all. That was my point.
Absolute compute power of GCN is much higher than kepler though due to the reasons given above. But it has yet to show this advantage in games.
No mate, i haven't taken you in a wrong way at all. We need members like you who can put facts owing to their experience. Everybody is free to give their points as per their will.
Read this:
*streamcomputing.eu/blog/2011-06-22/opencl-vs-cuda-misconceptions/
Although its an old article, it gives more insight. This will throw light to the fact that why i spoke about CUDA model and its libraries.
It will also tell why CUDA is used in supercomputer applications more than open-cl. Open-cl has a lot of catching up to do as far as providing libraries is concerned. Cuda's math libraries are vast and thus finds use in supercomputers.
Last edited: