Nvidia GTX300 series Supposed Benchmarks....and what it could mean for ATI's 5800's

[20mmrain]

3 GPUs on a single card and not a CAD program to use them. I doubt they'd spend the money to R&D that trainwreck.

Touche my good friend....... But there was nothing for Xfire at first either Nothing would ever work with out first a dream to do it.

They slap a 400W cooler on the 5970, and that is why it is so long

Yeah speaking of that since every Nvidia card I have ever owned has run hotter than ATI's (maybe just my experience) I wonder how hot that bad boy of Femi is going to run?

60C idle and 90C to 100C under load LOL

 

[KainXS]

the GTX380 will without a doubt be faster than the HD5870 but the question is will ATI release the HD6870 next year which will use the same architecture as the GTX380(mimd based) and stomp it into the ground, now this is news today, AMD is going to aim at releasing a Globalfoundries for GPU and CPU production, bypassing TSMC, and there are rumors that there will be no HD6870 released next year now, but there probably not true, since actual development on the 6K's are locked in specification.

 

[Binge]

Touche my good friend....... But there was nothing for Xfire at first either Nothing would ever work with out first a dream to do it.

There still isn't a multi-gpu CAD program that is used in the industry. ATI loses big bucks just to save face against NV with these dual GPU solutions.

Yeah speaking of that since every Nvidia card I have ever owned has run hotter than ATI's (maybe just my experience) I wonder how hot that bad boy of Femi is going to run?

60C idle and 90C to 100C under load LOL

Funny this 5870 I have in my rig right now is the first ATI card, besides the 5850, that's been more cool than 55nm nVidia products. I'm speaking in terms of degree C per FPS.

 

[20mmrain]

the GTX380 will without a doubt be faster than the HD5870 but the question is will ATI release the HD6870 next year which will use the same architecture as the GTX380(mimd based) and stomp it into the ground, now this is news today, AMD is going to aim at releasing a Globalfoundries for GPU and CPU production, bypassing TSMC, and there are rumors that there will be no HD6870 released next year now, but there probably not true.

Where did you see this not that I doubt you....

Funny this 5870 I have in my rig right now is the first card, besides the 5850, that's been more cool than 55nm nVidia products.

That's why I said maybe it's me..... Because every Nvidia card I have ever owned has been hotter than my ATI cards..... Hence 8800GT vs. my 3850 and GTX 385 FTW vs. 4890. Both of those cards at least for me were way cooler. That might of been your experience but that has never been mine.

 

[Binge]

I was just confirming that it's you.

 

[KainXS]

next gen discrete graphics for ati(HD6K) is rumored to come in 2011(this could mean Nov2010-Jan 2011 though)

http://www.brightsideofnews.com/new...h-globalfoundries-gpu-production-in-2011.aspx

but the cards could also be delayed for development on the GPU for the Xbox 3

 

[20mmrain]

I was just confirming that it's you.

I don't want to get on this too much because it is off topic...... But from some of the posts I checked I'm not the only one who thinks that But I'm sure if you look around it would probably even it's self out...... Some individual cards run hotter than others.

But just from my experience it's been that way.

next gen discrete graphics for ati(HD6K) is rumored to come in 2011(this could mean Nov2010-Jan 2011 though)

Ahh CC I have seen that but I called that Two years really maybe it should be looked at in a year and a half. 2009 to 2011.

But it could be brought out sooner if they need too I think you are right around Nov 2010

But that brings me back to my previous point ..... "possible Nvidia is making these to compete with the 6800 series in mind."

 

[jjFarking]

I think that the more closely you look at the actual figures, the more fake they have to be.
Just on one figure alone, if, on paper, the 'GTX380' looks like a beast, how come that same paper then kind of contradicts itself to being a 1.74 teraflops card, when the hd5870 is already 2.72 teraflops (the 5970 is 5+ teraflops)?
Also, with a fairly mild OC, the hd5850 will outdo a 'GTX360'.. Not exactly comforting stats for nVidia when you think about it.
Additionally, there's going to be an issue that ATI already came across with the 5970 - the ATX standard. Once you go over 300w, you go beyond that standard. I am unaware of any change happening soon to that standard, unless someone else here has more info?

I think someone had their hand(s) on their genitals when posting those stats..

 

[Lionheart]

yeah true about ATI's cards having more teraflops than Nvidias cards but i just thought it was different architecture and driver support!

 

[Fishymachine]

I think that the more closely you look at the actual figures, the more fake they have to be.
Just on one figure alone, if, on paper, the 'GTX380' looks like a beast, how come that same paper then kind of contradicts itself to being a 1.74 teraflops card, when the hd5870 is already 2.72 teraflops (the 5970 is 5+ teraflops)?
Also, with a fairly mild OC, the hd5850 will outdo a 'GTX360'.. Not exactly comforting stats for nVidia when you think about it.
Additionally, there's going to be an issue that ATI already came across with the 5970 - the ATX standard. Once you go over 300w, you go beyond that standard. I am unaware of any change happening soon to that standard, unless someone else here has more info?

I think someone had their hand(s) on their genitals when posting those stats..

even @1000 a 5870 might lag even on neutral ground (unless Cat 9.12/10.1 unlocks the idle 160 SP) but still +1

 

[Lionheart]

even @1000 a 5870 might lag even on neutral ground (unless Cat 9.12/10.1 unlocks the idle 160 SP) but still +1

Theres a idle 160 SP

 

[Jstn7477]

Theres a idle 160 SP

I think he might have meant 5850. (1440) + 160 = 1600 (5870.)

 

[NinkobEi]

edit. sigh, not needed.

 

[Nick89]

It means nothing these are fabricated by nvidia. I don't believe anything until wiz benchmarks it.

 

[Benetanegia]

Just on one figure alone, if, on paper, the 'GTX380' looks like a beast, how come that same paper then kind of contradicts itself to being a 1.74 teraflops card, when the hd5870 is already 2.72 teraflops (the 5970 is 5+ teraflops)?
Also, with a fairly mild OC, the hd5850 will outdo a 'GTX360'.. Not exactly comforting stats for nVidia when you think about it.

Raw peak GFlops means nothing. Actual sustained Gflops is what matters.

Here you can see actual Gflop numbers from SGEMM matrix multiplication benchmark: http://www.brightsideofnews.com/new...-1tflops---27x-faster-than-nvidia-gt200!.aspx

But as of SC09, the top five performing products for SGEMM 4K x 4K are as follows [do note that multi-GPU products are excluded as they don't run SGEMM]:
1. Intel Larrabee [LRB, 45nm] - 1006 GFLOPS
2. EVGA GeForce GTX 285 FTW - 425 GFLOPS
3. nVidia Tesla C1060 [GT200, 65nm] - 370 GFLOPS
4. AMD FireStream 9270 [RV770, 55nm] - 300 GFLOPS
5. IBM PowerXCell 8i [Cell, 65nm] - 164 GFLOPS

Firestream RV770, advertised 1200 Gflops, actual performance 300 Gflops, 300/1200 = 0.25
Tesla C1060 GT200, advertised 933 Gflops with dual issue and 622 gflops without, actual 370 Gflops, 370/622 = 0.6
GTX285 GT200, advertised 1062/708, actual 425 Gflops, 425/708 = 0.6

Next generation, projected performance on SGEMM:

RV870 --> 2.72 Gflops * 0.25 = 680 Gflops real
GT300 --> 1.74 Gflops * 0.6 = 1044 Gflops real

 

[yogurt_21]

Raw peak GFlops means nothing. Actual sustained Gflops is what matters.

Here you can see actual Gflop numbers from SGEMM matrix multiplication benchmark: http://www.brightsideofnews.com/new...-1tflops---27x-faster-than-nvidia-gt200!.aspx



Firestream RV770, advertised 1200 Gflops, actual performance 300 Gflops, 300/1200 = 0.25
Tesla C1060 GT200, advertised 933 Gflops with dual issue and 622 gflops without, actual 370 Gflops, 370/622 = 0.6
GTX285 GT200, advertised 1062/708, actual 425 Gflops, 425/708 = 0.6

Next generation, projected performance on SGEMM:

RV870 --> 2.72 Gflops * 0.25 = 680 Gflops real
GT300 --> 1.74 Gflops * 0.6 = 1044 Gflops real

numbers on dual precision aren't being backed up by even the most basic of google searches.

http://en.wikipedia.org/wiki/Nvidia_Tesla

http://www.engadget.com/2008/06/16/amds-firestream-9250-first-processor-to-top-1-teraflop/

on single precision nvidia advertizes that 933 number you list yet you randomly divided by a number I see nowhere else on the net. Again for ati on single precision they list 1200gflops so as an apples to apples comparison those are the numbers to use. That data is looking like an nvdia bias where none was needed, nvidia was already faster.

370/933 = .396
300/1200 = .25

much better ratio than ati's .25 (though I'm impressed they were able to maintain 25% of advertized on raw computing when they can only utilize 20% of the advertized stream processors at one time)

So to do this ridiculous assumption without a random bias dataset that appears nowhere else on the net.

RV870 --> 2.72 Gflops * 0.25 = 680 Gflops real
GT300 --> 1.74 Gflops * 0.396 = 689.04 Gflops real

again we must indicate that this is single precision, on dual precision the gt200 advertized 78gflops whereas the rv770 advertised 200 gflops. If anyone has a dual precision bench, please share it.

 

[douglatins]

Fake FAKE FaKe.... Bad graphs bad

 

[KainXS]

one thing about the SP's on Radeon cards is that there is ongoing debate if whether the FLOPS are calculated correctly, I for one think its flawed but thats me, I personally think ATI needs to quit with the whole, oh we have 800 or 1600 shaders, im not saying its not true but theres more accurate ways to say it than that.

 

[Binge]

Fake FAKE FaKe.... Bad graphs bad

Thanks for the stunningly original statement! Man I can do this all day. Correct me if I'm wrong, but haven't we covered the possibility from the moment the graphs were posted?

 

[Arctucas]

http://www.nordichardware.com/news,10412.html

 

[HalfAHertz]

It's all about programing. If your software can utilize the low level hardware, then you can use all of the SPs and hit those 2,7 Gflops. There was some guy who reached 1 Gflop with the HD4870 a while back ( or more precisely half a 4870x2 )

Edit: Found it http://forum.beyond3d.com/showthread.php?t=54842

 

[phanbuey]

It's all about programing. If your software can utilize the low level hardware, then you can use all of the SPs and hit those 2,7 Gflops. There was some guy who reached 1 Gflop with the HD4870 a while back

Edit: Found it http://forum.beyond3d.com/showthread.php?t=54842

thats what I thought... SGEMM is just an algorithm, and how you hit that tflop rating is up to the software...

HPC labs are always tweaking their software to get the fastest speeds, and they are doing the same matrix operations as the bench.

thanks for confirming

 

[troyrae360]

It's all about programing. If your software can utilize the low level hardware, then you can use all of the SPs and hit those 2,7 Gflops. There was some guy who reached 1 Gflop with the HD4870 a while back

Edit: Found it http://forum.beyond3d.com/showthread.php?t=54842

How do you measure your Gflops?

 

[zithe]

http://www.fudzilla.com/content/view/16843/1/

If fudzilla says that it's fake, it's got to be fake, right?

 

[Benetanegia]

It's all about programing. If your software can utilize the low level hardware, then you can use all of the SPs and hit those 2,7 Gflops. There was some guy who reached 1 Gflop with the HD4870 a while back ( or more precisely half a 4870x2 )

Edit: Found it http://forum.beyond3d.com/showthread.php?t=54842

thats what I thought... SGEMM is just an algorithm, and how you hit that tflop rating is up to the software...

HPC labs are always tweaking their software to get the fastest speeds, and they are doing the same matrix operations as the bench.

thanks for confirming

Yeah, but those are unrealistic. You can always find a benchmark that will make a higher use of the resources of a single architecture, pure matrix multiplication is one of the kind, but that is never going to be the case with real applications. The guy on the link is trying to find the best scenario for 4096x4096 matrix-multiplication on the Ati cards (finding the best block size, etc.), get out of that* and performance decreases dramatically. SGEMM tries to represent the performance that you would find in real applications.

*Only using different block size that could be required so that the code is portable to other GPUs or even the CPU in a setup with Ati cards, decreases performance dramatically, as you can see in the link. Add in the fact that different matrix size will benefit from different block size and add in some linear code or some transcendental functions and performance drops dramatically in the case of Ati.

Also the kind of optimization the guy is making there, goes absolutely against the purpose of API's like OpenCL. What's the point of making your code in OpenCL (because it runs in any platform) if you are going to make a different executable for every architecture? In that case you might want to port your code to the native API of each architecture and have the benefits aswell.