Kepler vs GCN: Which is the better architecture

[Jaydip]

Originally Posted by Red Hawk

Nothing you said is related to you saying "28 nm" as if it countered what another user said.



What? Both sets are contemporary architectures made on the same processing node. They have different advantages and disadvantages, but that doesn't make them not comparable.



Ok.



No comment.



And software has what to do with the architecture?



No cop-outs. Nvidia made GK104 available so that's what they will be tested by.

I mentioned 28 nm regarding the die size of Fermi. GK104 is not full Kepler, may be you should create this thread once GK110 for consumer cards come out or do you include Tesla cards as well? It means architecture is irrelevant unless it equates to better performing products.

[Haserath]

Originally Posted by VulgarDisplay

I don't think that has as much to do with GPU architecture as it does with the company's management decisions in the past. AMD was ridiculously mismanaged in the past 5 years and they have been paying for it. If they had moved into HPC with nvidia things may not be as swayed to Nvidia in that area as they are now.

Except AMD still doesn't seem to have anything comparable to Nvidia for HPC. Theoretical numbers may sound good, but that's nothing if they can't actually get them.

Kepler is better for mobile, Kepler competes on the desktop, and Kepler still demolishes HPC.

The one place GCN has a haven atm is descrete desktop with good deals and more customization(for lack of a better word).

[Silverforce11]

Bitcoin domination, and the news that Bitcoins are now accepted by international banks like just paypal is a pretty big deal for the short term.

The long term for AMD's compute? Well, Tahiti being in one of the most efficient super computers isn't too bad. They have good hardware, always has, just need great software support to help push their HPC advance. Its certainly not an easy market to penetrate with NV's dominance though. Intel isn't sitting idle also. Realistically, AMD has no chance in GPU-HPC.

As for GCN vs Kepler for gaming, 7970 Ghz > 680, with a smaller power (and die size) delta than gtx480 vs 5870. On perf/w, GCN is 20% less efficient for the top models and more efficient for mid-range and lower. The only logical conclusion is that its an excellent architecture that does not compromise on compute while being fastest for gaming, at the expense of very little headroom (die space & power use). Kepler is a better architecture for NV making $$ certainly, great gaming perf, relatively tiny die space, ka-ching!

Both are great IMO. GK110 vs gk104 in gaming will be interesting to see, i suspect it will look awful in perf/w and perf/mm, but thats to be expected since its not its primary purpose.

[Red Hawk]

Originally Posted by Jaydip

I mentioned 28 nm regarding the die size of Fermi. GK104 is not full Kepler, may be you should create this thread once GK110 for consumer cards come out or do you include Tesla cards as well? It means architecture is irrelevant unless it equates to better performing products.

Ok? No one is directly comparing Fermi's die size to GCN's die size; we are comparing the die size ratio between Evergreen and Fermi against the ratio between Kepler and GCN. The ratio between Evergreen and Fermi would be the same at 40 nm or 28 nm.

"Full Kepler" may not be very available but that doesn't mean we can't evaluate the architecture. And if you don't want to talk about architecture, then you're free to leave this thread, because that's what it is about.

[Jaydip]

Originally Posted by Red Hawk

Ok? No one is directly comparing Fermi's die size to GCN's die size; we are comparing the die size ratio between Evergreen and Fermi against the ratio between Kepler and GCN. The ratio between Evergreen and Fermi would be the same at 40 nm or 28 nm.

"Full Kepler" may not be very available but that doesn't mean we can't evaluate the architecture. And if you don't want to talk about architecture, then you're free to leave this thread, because that's what it is about.

Really let me know how I'm open to such grand ideas.

[Red Hawk]

Originally Posted by Jaydip

Really let me know how I'm open to such grand ideas.

Are you really thick enough that you need me to spell it out for you? Or are you just trolling? Please say trolling, I prefer to think of people as being intelligent; trolls, while often rude, can still be intelligent. (the best trolls are often quite intelligent in their trolling).

[Jaydip]

Originally Posted by Red Hawk

Are you really thick enough that you need me to spell it out for you? Or are you just trolling? Please say trolling, I prefer to think of people as being intelligent; trolls, while often rude, can still be intelligent. (the best trolls are often quite intelligent in their trolling).

This concludes the thread.You have absolutely no clue about gpu architecture so stop creating this non sense threads.If a person say he can evaluate a architecture without any relevant data that speaks for itself.

[Red Hawk]

Originally Posted by Jaydip

This concludes the thread.You have absolutely no clue about gpu architecture so stop creating this non sense threads.If a person say he can evaluate a architecture without any relevant data that speaks for itself.

Did I admit to not having relevant data? There's relevant data all over the Web -- benchmarks testing gaming speed, power draw, heat levels, general compute power etc. There's also public information about how large the dice (is that the proper plural form for that word...?) are and how many transistors are in them. All of this can be used to compare the architectures. Now stop acting like you can tell me what to do.

[Jaydip]

Originally Posted by Red Hawk

Did I admit to not having relevant data? There's relevant data all over the Web -- benchmarks testing gaming speed, power draw, heat levels, general compute power etc. There's also public information about how large the dice (is that the proper plural form for that word...?) are and how many transistors are in them. All of this can be used to compare the architectures. Now stop acting like you can tell me what to do.

Lets try it for the last time.Kepler doesn't mean only GK104 but GK110 as well.So far we only have theoretical data about GK110 but no actual benchmark data.If you change your thread title to GK104 it would be more accurate.

[n0x1ous]

Originally Posted by Will Robinson

mm...good idea...let's compare some mythical,non existent chip with AMD's freely available production chip...that'll work...

Mythical and non-existent? I think its pretty well known that HPC customers have received and are actively using GK110 Tesla cards.

Nice try

[Red Hawk]

Originally Posted by Jaydip

Lets try it for the last time.Kepler doesn't mean only GK104 but GK110 as well.So far we only have theoretical data about GK110 but no actual benchmark data.If you change your thread title to GK104 it would be more accurate.

The fact that little information about GK110 is available doesn't somehow invalidate information from GK104, 106, and 107. We can certainly compare the architectures just based on that information; might we have to change the conclusions from that comparison upon receiving more information on GK110? Maybe; then again, maybe not.

Whether or not it is possible to compare architectures is a different question than whether or not it is productive or useful to compare architectures. You seem to be taking (rather shaky) reasons against the latter and acting like they also are reasons against the former because they are against the latter.

[Elfear]

Originally Posted by Jaydip

Lets try it for the last time.Kepler doesn't mean only GK104 but GK110 as well.So far we only have theoretical data about GK110 but no actual benchmark data.If you change your thread title to GK104 it would be more accurate.

So we need to wait until both next gen cards launch before we can compare the two architectures?

[AnandThenMan]

Originally Posted by Jaydip

Lets try it for the last time.Kepler doesn't mean only GK104 but GK110 as well.So far we only have theoretical data about GK110 but no actual benchmark data.If you change your thread title to GK104 it would be more accurate.

You want to compare theoretical vs. what is actually available?

[beginner99]

Don't really care but AMD (GCN) offers way, way better $/performance at least were i live. I would have to be plain stupid to by NV. Same as with CPU but there AMD is the clear looser.

[96Firebird]

Originally Posted by beginner99

Don't really care but AMD (GCN) offers way, way better $/performance at least were i live. I would have to be plain stupid to by NV. Same as with CPU but there AMD is the clear looser.

Right. This thread is about architecture, where the price of the card really shouldn't come into play.

[3DVagabond]

Originally Posted by beginner99

Don't really care but AMD (GCN) offers way, way better $/performance at least were i live. I would have to be plain stupid to by NV. Same as with CPU but there AMD is the clear looser.

That's the same here, but for consistency, it's easier to talk North American market. But I agree, anyone buying nVidia here you would have to be misinformed or a fanboi, because pricing has never been in line with performance between Kepler and GCN.

[AnandThenMan]

Show me GK110 benches, and we'll consider it part of this discussion.

[96Firebird]

Originally Posted by AnandThenMan

So show me GK110 benches, and we'll consider it part of this discussion.

Why are you asking me this?

FWIW, I went and looked for some (using keywords GK110, Tesla K20, and Quadro K5000). Couldn't find anything just yet, but benchmarks for professional cards are surprisingly hard to come by. TH had some for the Firepro W9000, a quick browse showed it beating the Quadro 6000 in some programs, and losing in others. This is the W9000 preview from Anandtech, but I couldn't find the part 2 they promised. Perhaps they are waiting to get their hands on GK110?

Edit - Forgot the TH link...

[AnandThenMan]

Originally Posted by 96Firebird

Why are you asking me this?

Multi-quote glitch, or my error. Sorry.

[notty22]

Here is some experiences with the K20.
http://blogs.nvidia.com/2012/12/test...t-to-research/
It seems Nvidia has a test drive program that allows, programmers/teams to log in to a K20 remote cluster and test.
www.nvidia.com/GPUTestDrive