Official Statement from Otoy towards GTX 980/970

[Vue2Octane]

Great...can not compare cuda cores side by side from different architectures!?
That's just what I needed to hear.
Then why even putting it as a number in every nvidia card spec without any 'architecture factor'? Why putting it on the OTOY website as the single one important number for Octane performance. If I go tommorow and buy a Card for Octane, then I look at the Otoy website, I see cuda cores means faster, i go to nvidia website picking the one with the most cuda cores. Period. Later the guys will go like 'sorry, it's so slow because it is Newton Cuda not Davinci Cuda' and 'sorry that we did not think of specifying a cuda core by using any performance standard that is independent of the architecture'. And I will go "Thank you very much, I will rely on forum entries in the future'

[glimpse]

OTOY is a company & they use their time & money of investors (and those who buy software) to make software..they are providing software. That's their work.

Yes they have a community, unlike RandomControl, who have simillar product - but no forum at all! So if You have a possibility, simply do Your homework, ask someOne around.

now nvidia will say: GTX cards are not meant for Pro work - buy our Tesla & Quadro. These products are marketed not by they sluggish performance in games, but for compute possibilities (especially tesla line). They have numbers DP & SP. For OctaneRender SP is what matter - the higher the number, the faster the card does calculation. Other things to keep in mind is memory (or vRAM).

That's it, it's not a rocket science. But You save ton of money using gamer cards instead of pro cards so don't make too much noise 'cos You're lazy to do some homework, when You have all the possibilities.

Yes, they can do better (infact we all can do better in certain areas of our life),..but You have such a great forum & if You don't have knowledge don't be so arrogant to ask questions..-that's nothing bad about it.

Now OTOY is not responsible for all the changes nVidia choose to make..Though I do agree with You that some information might be explained way better - but just imagine how long that explanation would be & if it be too long, majority is not going to read it at all.

P.S. Do You blame Chaos Group that they don't write on their vary website that AMD's cores are not on par compared to INTEL cores? No because it's not their business! You go on shop and ask, if in the shop seller doesn't know You get online & read some reviews..-that's technology, it changes too fast..

[Vue2Octane]

Glimpse, that's all true. There is a nice interview with J. Urbach and a nvidia guy online outlining the differences between gtx and the pro families.

But for Cuda and how it is defined I think you misunderstand something. Firstly, CUDA is a programming API. There are no actual phyiscal cores there, each with say 5 transistors. Pity there is no definition to be found. A Cuda core as a virtual unit has to be something like 'that and that many well defined operations in that and that time' - it has to be some kind of performance standard. Otherwise nVidia could not just put 'cuda cores' and then the number in every card spec disregarding the architecture. It might be something like 1 Cuda Core = 1000 operations/ms. Then Gtx780 makes say 2300000 operations/ms and therefore has 2300 cuda cores and Titan-whatnot makes 10000000 and has 10000 cuda cores. Random numbers to illustrate, but it has to be some generalized performance standard, not an actual core. I am sure! But then I was wrong before, when I was sure that I was sure. ... oh, scene compiling finished.

[glimpse]

Vue2Octane, TFLOP (1 Teraflop = 1,000,000,000,000 (Trillion) operations per second in floating points) is that unit that defines the power of GPU. SP performance is what matter for OctaneRender.

Now nvidia target GTX line to gamers, where other driver optimisations matter more than raw performance..it is somehow proportional, but that's not the only factor..as PRO line might have similar SP performance but would behave drastically lower on games, 'cos of absence of optimised drivers.

PRO line is targeted to professional..it has different driver & different way of marketing, where those numbers are actually marketed, 'cos that's what matter for potential buyers (they don't care how many FPS it does in one or the other game..'cos they don't use these products in such environments).

For gaming cards is opposite story, as gammers only matter for FPS (frames per second) in their preferred title..they don't care, bout cuds, SP or DP.. if 680 gives better FPS then 580 for the same price, draw less energy then previous model they don't care about anything else (well maybe the price). & You're in that market to look for the card that is not meant by nvidia to be used this way =)

it seems they are putting Titan Black, Titan Z into entirely different category where they start to market those as entry level PRO cards for compute applications (even if they are still under GTX line). but they know how strong are they & how some people use them..-for this reason we see elevated price..

so back on our topic..don't expect for entry level family sedan that seller to will give You 0-100km/h numbers in seconds - it's not their ace, when at the same time nobody writes how much l/km eats high end Ducati Panigale - 'cos nobody cares.. - You don't drive moto for saving fuel..

the same goes here..GTX line is nor marketed for insane performance: benchmark sites make cases (& then comes nvidia to tomsHardware & ask to pull of some charts 'cos they miss represent their ideas..-too much OCed card in different loads that it was meant for)..FPS for games is what matters..

[Vue2Octane]

Yes, that's true. We use some Quadros for those reasons at work. Also servicing etc.
Teraflops as gpu standard is also ok and clear, but i am still wondering how a cuda core is actually defined. It seems a bit like a magic unit nowhere to be found. Memory bandwidth etc all have clear units that can be traced back to existing standards (e.g. what is a byte-> well known), but 'cuda core' does not. It makes it hard to eventually go into indepth performance analysis prediction if there are such 'gray area' specs out there such as the 'cuda core'.

[Seekerfinder]

I agree with Vue2Octane. When the technology was brand new, cuda core count was the most sensible gauge. But then we had the Fermi-Kepler change which confused a lot of people. Since Octane's speed is one of its big selling points over CPU renderer's, it would be great to get a common standard that could be easily be compared.

Here is my suggestion: samples per second for a single dedicated (non-monitor) card by type (one for Titan, one for a GTX 780 etc...) based on one of the sample benchmark scenes for DL, PT and PMC respectively. Every time there is a new release, we do a new thread to update the stats. That way, we can compare not only cards to one another, but also releases. This could be a community effort we can all contribute to (just disable the other cards for the test).

What say ye?

Seeker

[glimpse]

Here is my suggestion: samples per second for a single dedicated (non-monitor) card by type (one for Titan, one for a GTX 780 etc...) based on one of the sample benchmark scenes for DL, PT and PMC respectively. Every time there is a new release, we do a new thread to update the stats. That way, we can compare not only cards to one another, but also releases. This could be a community effort we can all contribute to (just disable the other cards for the test).

What say ye?

Seeker

I'd say wonderful idea! Test same bench scene after a release on one card (GPUz) on top of running Octanerender printscreen would be nice addition as we would know all the metrics too! let's start it, Seeker, with next release =)

[Vue2Octane]

+1 seeker.

But I would prefer the Octane Team to release the benchmarks for each card. If the forum members do it, the thread will derail in no-time and will be mess.

[Seekerfinder]

But I would prefer the Octane Team to release the benchmarks for each card.

Aint gonna happen... Firstly, because they would have purchase at least one of every new card out there. Secondly, they do not have the time. Besides, I would prefer a small team focusing on what they do best.

If the forum members do it, the thread will derail in no-time and will be mess.

We just need to manage it to keep it 'pure' with a heading with some base rules. Most guys here are decent enough. I think we can do it. Once in a while some kind soul could throw the numbers on a chart for an overview.

Best,
Seeker

[glimpse]

agree with Seeker, pluss there are a lot of same model cards in a different versions - with such user base we could have plenty of info. advice from my side would be moderate these threads a bit, so they wouldn't have any comments, just plain info (screenShot, card model & numbers). In the first post we can do a short conclusion with those numbers.