AMD: R7xx Speculation

Status
Not open for further replies.
I was going to say, if the P test were mostly texture-bound, would we see a 3850 equivalent to a 9600GT? I thought cheaper cards tended toward more math than texture power.

I forget the math-texture ratios for RV630 and G84, so I don't know what to make of the 8600s spanking the 3650 (aka 2600, I believe).

The Vantage whitepaper doesn't help someone clueless like me to interpret what part of the GPU's getting stressed. It looks like GPU test 1 stresses math ("aniso shaders," cloth shaders) and test 2 features instancing and POM (branching?). I'm puzzled why the 3650 edges the 8600 in the POM feature test (maybe just the latter hitting a FLOPs wall), and have no idea how cascading or variance shadow maps pressure a GPU.
 
HD3850 should have way more math than 9600GT:

http://www.firingsquad.com/hardware/3dmark_vantage_performance/page9.asp

and I think GPU test 1 has more simulation (maths) than GPU test 2.

So, erm, it doesn't look like P is primarily math bound. Or, if it is, there's a huge shortfall in math efficiency in RV670 for these tests.

Jawed

I think it is math-bound. We know R6xx's ALU utilization is lower than G8x/G9x's, particularly in games. I think 3dmark 06 is about the only 3d app. that shows ALU utilization rates anywhere near 100% on R6xx hardware.
 
Here's a useful summary:

http://www.xbitlabs.com/misc/picture/?src=/images/video/msi-hd3650/spec.png&1=1

HD3650 as a percentage of HD3850:
  • bilinear 54%
  • GFLOPs 41%
  • fillrate 27%
  • GB/s 48%
  • test 1 14%
  • test 2 33%
So test 2 could be math bound on ATI. Sadly with test 1 there's either something horribly broken in the driver or there's a dark secret to RV635. Dynamic branching should be better on RV635 (smaller batches than RV670)...


8600GTS as a percentage of 9600GT:
  • bilinear 52%
  • GFLOPs 45%
  • fillrate 52%
  • GB/s 56%
  • test 1 44%
  • test 2 47%
which seems to imply that both tests are math bound on NVidia. With the very low ALU:TEX ratio of this NVidia hardware, that shouldn't be a surprise.

Jawed
 
Last edited by a moderator:
Right, I was thinking about IGPs, which aren't really worth discussing, anyway. Thanks for the summary.

As for ATI and GT1, look at the Feature Test scores on p.10 of FS' article and you'll see the 3650 fares just as badly with the cloth shader (GT1) as with POM (GT2). Do cloth shaders stress branching or fetches (like the water surface shaders using VTFs in Pacific Fighters)?

I remembered digit-life had some parallax/offset mapping tests, but their results show a RV630 basically equal to a G84, quite different from what Vantage shows with their POM test. Either POM is too far removed from d-l's method to warrant comparison, or NV's improved performance with some driver work. Obviously if I don't even know the difference (if any) b/w parallax/offset mapping and POM, I should post less and read more, and there's a wealth of info in those first few links. A brief look shows POM to be different than parallax mapping, so that would explain the performance differences b/w the two methods (d-l's steep parallax PS3.0 method is "heavy" on the ALUs, fetches, and branching, but Tatarchuk's "Practical POM" paper and the benchmarks above indicate different bottlenecks b/w parallax and POM).
 
Taking the other tests.

ATI:
  • Texture - 39%
  • Colour - 41%
  • POM - 18%
  • Cloth - 29%
  • Particles - 50%
  • Perlin - 25%
NVidia:
  • Texture - 52%
  • Colour - 52%
  • POM - 50%
  • Cloth - 54%
  • Particles - 42%
  • Perlin - 50%
On NVidia Cloth looks like it could be bandwidth bound and Particles look like it could be math bound.

ATI's a mess. Prolly a driver mess. Waste of time kinda mess.

Jawed
 
Do cloth shaders stress branching or fetches (like the water surface shaders using VTFs in Pacific Fighters)?
Branching should improve on RV635 in comparison with RV670 as the batches are half the size.

Cloth gives the impression of being bandwidth bound on NVidia which is puzzling. That might be caused by a lot of incoherent texture fetches (i.e. cache can't keep up so bandwidth is hit hard). Someone needs to get a hold of the shader code :p

Jawed
 
How do you conclude that HD3650 has 9% more fillrate than HD3850?
:oops: Sigh I thought I'd wheedled-out my errors - similar reason for my GFLOPs slip-up.

Might as well add that it makes test 2 look like it's tending towards being fillrate limited on ATI (variance shadow mapping, writing the shadows?).

Jawed
 
Cinema 2.0

http://translate.google.com/transla...04578/20080604037/&sl=ja&tl=en&hl=ja&ie=UTF-8

June 4, 2008, AMD COMPUTEX TAIPEI 2008 is carried out according
to a new generation notebook PC platform "Puma" and the unveiling of the
next generation GPU "RV770" Sneak Praevius were conducted.

Unfortunately, the card itself is not disclosed to the public, ATI Radeon
for the new technology is introduced at the demonstration, so that we
would like to hasten the movie.

Suddenly that the tripod is not available because the quality of'd love
to be relentless, because one is, AMD is the next generation GPU, the
proposed "Cinema 2.0" (※ movie-like graphics and realistic You have one)
demonstrations, two for the GPU-ray tracing demonstration.
 
http://translate.google.com/transla...04578/20080604037/&sl=ja&tl=en&hl=ja&ie=UTF-8

June 4, 2008, AMD COMPUTEX TAIPEI 2008 is carried out according
to a new generation notebook PC platform "Puma" and the unveiling of the
next generation GPU "RV770" Sneak Praevius were conducted.

Unfortunately, the card itself is not disclosed to the public, ATI Radeon
for the new technology is introduced at the demonstration, so that we
would like to hasten the movie.

Suddenly that the tripod is not available because the quality of'd love
to be relentless, because one is, AMD is the next generation GPU, the
proposed "Cinema 2.0" (※ movie-like graphics and realistic You have one)
demonstrations, two for the GPU-ray tracing demonstration.

Holy Crap! I know that is a tiny little, off-screen camera, low-res video... but that looked like a real city! @o@/ I'm really looking forward to seeing that demo!
 
Taken from ITOCP, read with salt. Pepper recommended :D


4850 Vantage P5847, X2609

On the same platform, a 9800GTX does P5816, X2104

If that's any indicator of AA, AF, and 4870 performance... ATI probably has improved a lot on their weaker points but on pure shader math, lesser stuff is done. This probably suggests a much better balanced chip, if slightly better suited for dual GPU usage.

Again, anticipating the STALKER performance.
 
mm... CJ said that R700 scores about 5500 points in Vantage Extreme Test.
It's more than the double of HD3870X2 performance. Then ,if they haven't improvent the scaling, this would mean that 4870 could have about twice the performance of a 3870. But if 4870 double the 3870, will be also about more then 50% faster then 9800GXT... and a bit faster then 9800GX2 in some test..
 
Perhaps you might be interested in this. High-res photos of the Gigabyte Radeon HD 4850, taken just hours ago at Computex. You can see them exclusively here (don't mind the text, no new discovery there). Notice the card isn't actually Gigabyte, but an AMD Engineering Sample made in Canada.
 
ECS/PCPartner will be the ref manufacturing partner for these, IIRC. Other AIBs will intro their designs later. I hadn't really noticed the DIP switches & header on the back before.
 
It's very probable that PC Partner will take over mass production. I merely pointed out that this very card is made in Canada, in the small assembly facility in the former ATi HQ in Markham. By the way, the manufacturing date is 25th February 2008.
 
Status
Not open for further replies.
Back
Top