Nvidia: We Expected More from the Radeon HD 7900 Series

Don't worry guys these are pretty much stuck together from other rumours that have already been proven to be false.

So a big false chart made up of old false info :3

We will be hearing some official news from nvidia this month so I'm personally going to just wait for that before judging.

 

Cool, can't wait till they're released...

regards,
RagDoll.

 

I ain't choosing anything, take a look at nvidia forums 400/500 series section and driver section.

Which chips suffer the most? yep gf104/114.


anyway i agree with DesGaizu, lets wait and see how this gk104 turd turns out lol

 

GF104 did seemingly have an unusually high rate of incident.....but, that doesn't necessarily mean it was a bad GPU. They still performed very well for their market segment if you were lucky enough to get one that functioned properly.

 

Yeah its not a bad gpu, quite fast actually, but the flaws are just to big to ignore it.

And i have a really bad feeling its gonna be the same with GK104. I already said that to one nv tech guy trollo aka Rollo, nvidia better not make the same mistake as with GF104/114 shift everything on SW.. But he was to proud and kept bashing me, saying then i should buy 7970 lol

 

Hope they release it soon, one of my 470's is toast so I can't play NFS:TR on full I don't think, so waiting and gonna get a 660 when they are released... Here's hoping it's mid February and not end of Feb...

regards,
RagDoll.

 

There could be some truth to those statistics i mean,guess when you consider that going from GTX285 to GTX480 the sps doubled from 240 to 480, then it isn't too wild a claim that going from GTX580 to GTX680 they will double again (512->1024), similalry going from 8800GTX to 280GTX the shaders nearly doubled (128->240).

 

Thats true vet,even a busted clock might be right sometimes.

 

well they are dropping shader core clocks down to core clock and so no wonder if 512 to 1024.
and hmm else will see how things go, but still time till something concrete will come out.

 

Let's see nvidias crippled physx pull something like that and be smooth while doing it without any data copied back to memory.

OpenCL gpu rigidbody pipeline, 110k bodies in real-time on a Radeon 7970
http://www.youtube.com/watch?v=8jGZv1YYe2c&list=HL1328245816


more stuff from OpenCL
http://www.youtube.com/watch?v=bvjOXl4KUiM&feature=related

http://www.youtube.com/user/OpenCLStudio?feature=watch

 

Seems anything but smooth or optimised at the moment.

30fps on a 570 with only 4k bodies, and 50% GPU usage.

 

Well my own GTX 580 can run that demo just fine at 60 fps with the default 4,096 cubes (the framerate appears to be locked at 60 fps and I've no idea how to increase the cube count). The lack of memory might be an issue as the demo uses 983 MB with 4,096 cubes. The HD 7970 has 3 GB so it obviously can handle more than the 1.5 GB GTX 580. But there is a 3 GB GTX 580 too. Perhaps someone can run the test on it with 110,000 cubes?

I'm not sure what point you're making here though because the current NVIDIA and previous AMD cards can run this OpenCL demo just fine so long as they're using the latest drivers and OpenGL, which this demo also uses, tends to be faster and less buggy on NVIDIA hardware (Oddworld: Stranger's Wrath or RAGE anyone?). NVIDIA cards can do both CUDA-based PhysX and OpenCL whereas AMD cards can only do the latter.

 

Let's see this used in a game.

I mean seriously, tired of seeing theoretical physics examples. Show me this **** in a game running in OpenCL and I'll agree that it's more impressive then PhysX. Until then it's pointless. Developers don't care about how memory efficient systems are, they want something that's easy to implement and is attractive to people.

 

I never managed 60fps, kept dipping down to 30.

Ran it again, and seems ok, 60fps and 40% GPU usage this time, far too system heavy to use in a game for several years.
Is there any way of upping it from 4096, as i'm not seeing how at the moment.

 

not impressed :3eyes:

 

The batch files contain variables for x_dim, y_dim and z_dim. 48 for each gives 110,592 cubes. For some reason the NVIDIA one only generates 4,096 cubes but, oddly, the GTX 580 can run the NVIDIA and AMD tests.

I'm running the AMD one with interop now with 110,592 cubes and it is running smoothly albeit it starts at 9-10 fps. As the demo progresses though the framerate gets faster (it is currently at 15-20 fps as cubes disappear off-screen!). No doubt the Kepler cards will be able to run it smoother, as fast or faster than the HD 7970.

It's nothing impressive anyway. It's a showy tech demo that bears no resemblance to anything we're likely to see in a game.

 

Just worked it out now, must of been blind, cheers Darren

Yeah, far too system heavy for what your seeing, even if something like this did make it into a game this year, which we all know it won't, it would have to be scaled way way down to allow for the rest of the visuals.

There are a few impressive and optimised PhysX tech demos, but not like anything optimised has ever made it's way into a game.

 

^
That's over 100k independent objects what did you expect, to run at 100fps? you will need something stronger for that.




And you all seem to miss the main point namely efficiency, its without any unnecessary calls something physx lacks. Alice2, new batman or any hw physx game is a great example.


Anyway Physx 3.0 looks promising, but bullet is still better/more optimized in such tech demos.
Now if it were openCL and without any crippling code in it, it would be a lot better i have no doubt about that.

 

Memory is cheap....why worry about it? lol

Developers are only worried about whether it's possible to run it....

I'd like to see more than block demos myself.

 

Did Nvidia not get the Supersled demo running on a single 480 with a million though, with 100k being the default.

I've not saw any PhysX 3.0 tech demos to compare, so no idea if Bullet is better or not, it was certainly head and shoulders above any of the 2.8.x ones, but no one was/is willing to use it, so it unfortunately it never had a chance.

It looks like it will either be all CPU physics, or Nvidia GPU PhysX as no one else is putting any sort of serious effort into it.