GTX470 benchmarks / comparison

[Amplitude]

this video is a bit old, I think the scale of that parameter has changed since.

[GDRS]

Something is wrong here.

I get from 4,87 to 4,95 all the time.

With GTX 280.

It gets to 4000 s at the time of 3.28.

Look at your self.

[radiance]

Something is wrong here.

I get from 4,87 to 4,95 all the time.

With GTX 280.

It gets to 4000 s at the time of 3.28.

Look at your self.

switch the kernel to pathtracing

Radiance

[Amplitude]

you're using directlighting, switch to the pathtracing kernel.

[GDRS]

Ok we will do that.

Thanks in advance.

[Florinmocanu]

Hello there, my first post here on the forum.

Is there a version of this benchmark for 2.2? I have the latest RC3 release and i cannot open the benchmark, it gives me an invalid version error.

Also, i read some things about people asking why fermi is just 2-3 times faster than a 9800 GT (G92 with 112sps).


First thing, computing/rendering power of a video card counts on double precision power, that is the important stat we need to look at when considering a GPU.

GF100 chips (fermi) have around 1.1-1.2 Teraflops single precision and for 2 single precision operations they can do 1 double precision operation, so around 600-550 Gflops of DP power.

A G80/G92/ GT200 based card, for each 8 single precision operations can do 1 double precision operation. A GT 200 (GTX 285) has around 0.8-0.9 teraflops SP and i think around 90-100 Gflops of DP.

But, Geforce versions of fermi are neutered, the have only 164 Gflops double precision compared to some 500 Gflops of quadro/tesla cards based on fermi (which should have been the real power of normal Geforce as well). It was a decision made because of 2 reasons.

1. increase sales of quadro/tesla, because people would buy them for extra rendering/compute power + extra memory.

2. reduce power consumption on geforce products. Double precision doesn't count for gaming so they disabled some computing capabilities to minimize power consumption in order to be able to get decent clocks out of fermi chips without going over 300 w ( which is the limit PCI-E 2.0 standard limit).

Compared to 164 Gflops of a 480, a GTX 280/285 has around 80-90, so you can see why earlier products, G80/G92 and GT200 are still decent compared to Fermi.

I would personally advise, until DX 11 comes to Octane, to find a 2 GB GTX 285. I think it's the best combo of speed/memory available right now on the market. Well, until we see 3GB GTX 480, if the manufacturers will make such a design in the future.


Edit: Another thing, about ATI cards. And why i wish Open CL/Direct compute can find a place in octane in the future.

Ati cards are not neutered in any way, since they don't have tight power limitations, their around 180-190W for a 5870, so well inside the 300W limitation. A 5870 has 2.7 Teraflops single precision and 1/5 DP power, so around 540 Gflops DP. If Octane could use Ati cards, ati cards would rape everything. But, it seems AMD should involve a bit more into software development/support to promote computing on Ati cards.

Cheers,
Florin

[Amplitude]

Interesting reading, thank you !

Is there a version of this benchmark for 2.2?

yes, you have to download the new demosuite. Be sure to switch to pathtracing as it is not by default.

I would personally advise, until DX 11 comes to Octane, to find a 2 GB GTX 285. I think it's the best combo of speed/memory available right now on the market. Well, until we see 3GB GTX 480, if the manufacturers will make such a design in the future.

Exactly what I did, and it comes cheap if you can find them secondhand too.

[Florinmocanu]

Ah, got it to work.

My GTS 250 512 mb averages around 2.28 Mpx.

System specs are I7 920 OCed to 4.031 Ghz, 6 GB DDR3 1536 mhz and GTS 250 from Leadtek, has around 5% OC from factory. it's an OC version.

[Amplitude]

My GTS 250 512 mb averages around 2.28 Mpx.

good, but that is directlighting speed With pathtracing it should be around 0.84 Ms/s

[Florinmocanu]

Actually, it did just 0.58 Mpx.

But, why is path-tracing emphasized on? Gives more accurate/faster results? Or what?