Kepler vs GCN: Which is the better architecture

[parvadomus]

I think GCN is the better architecture overall, it has great compute capabilities, and excellent power efficiency.
A lot of people in this thread fail to understand that we are talking about architectures, not particular implementations like Tahiti or GK104. Comparing them is just stupid because Tahiti is a high-end part, oriented to game/compute, while GK104 is a mid-range totally oriented to gaming.
Lets talk about comparable implementations, like:
-GK106 vs Pitcairn [comparable sizes, and core clocks]
-Size: 214mm^2 vs 212mm^2
-Relative perf: 100% vs ~107%

-Efficiency: Pitcairn is 16% more efficient.

-Compute: better not to talk about this, Pitcairn just blows away even GK104.
This is the way I see it, if AMD decides to bring a gaming oriented die for high-end gaming for HD8000 series, then nvidia will have a HARD time to deal with it.

 

[96Firebird]

A lot of people in this thread fail to understand that we are talking about architectures, not particular implementations like Tahiti or GK104. Comparing them is just stupid because Tahiti is a high-end part, oriented to game/compute, while GK104 is a mid-range totally oriented to gaming.

So compare the architectures. Kepler includes GK104, GK106, GK110, as well as the mobile parts. GCN includes Tahiti, Pitcairn, as well as their mobile parts.

 

[wand3r3r]

OT but I wonder if there would be any poll between ATI and Nvidia on this forum that ATI wouldn't win? This place is skewed big time for big red

Until NV is the clear price/performance winner why should they be more favored? Many of us are swayed by performance and price or a ratio thereof. Please enlighten us in which area NV should be ahead of AMD where they aren't or haven't been in the polls?

 

[SirPauly]

What difference does it matter? As long as there is compelling choice for individuals to choose what best fits their subjective tastes and tolerances -- that is key to me.

 

[Red Hawk]

Adaptive Vsync is not a feature it's a marketing gimmick same with Physx and whats Cuda you mean Streaming Processors well AMD has them as well.

"Stream Processor" is an AMD term; you'll never see Nvidia use it in their product descriptions. Nvidia's counterpart to that term is CUDA Core. A more accurate general term is shader unit.

 

[Lepton87]

Because nobody is going to review one card. Start from 100 K20Xs and we talk.



Actually it's 4-5%. (680 vs 7970GHz)

http://tpucdn.com/reviews/Club_3D/HD_7870_jokerCard_Tahiti_LE/images/perfrel.gif[
[/QUOTE]

Who plays at all resolutions? Maybe show as an image at 640x480 and conclude that GTX680 is faster.
[img]http://tpucdn.com/reviews/Club_3D/HD_7870_jokerCard_Tahiti_LE/images/perfrel_2560.gif That's the most relevant resolution when it comes to high-end cards. 1080p might also be important somewhat.

 

[SirPauly]

I think GCN is the better architecture overall, it has great compute capabilities, and excellent power efficiency.
A lot of people in this thread fail to understand that we are talking about architectures, not particular implementations like Tahiti or GK104. Comparing them is just stupid because Tahiti is a high-end part, oriented to game/compute, while GK104 is a mid-range totally oriented to gaming.
Lets talk about comparable implementations, like:
-GK106 vs Pitcairn [comparable sizes, and core clocks]
-Size: 214mm^2 vs 212mm^2
-Relative perf: 100% vs ~107%

-Efficiency: Pitcairn is 16% more efficient.

-Compute: better not to talk about this, Pitcairn just blows away even GK104.
This is the way I see it, if AMD decides to bring a gaming oriented die for high-end gaming for HD8000 series, then nvidia will have a HARD time to deal with it.

How does one really decide? Gaming? Professional? Performance? Efficiency? Just discrete desktop? Discrete mobile?Total discrete? Everything GPU related?

How does one define better? Share? Revenue? Profits? Performance? Features? Efficiency?

For me it was about raw gaming performance from a desktop discrete high end GPU and feel with maturity, GCN is separating from Kepler.

 

[Silverforce11]

Fermi vs Evergreen parrallels anyone?

 

[Haserath]

Food for thought: What supercomputer do they power?

 

[VulgarDisplay]

Food for thought: What supercomputer do they power?

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.

 

[HurleyBird]

1) GK110 is already out there

And we hardly know anything about it's performance,

2) 40 nm vs 28 nm yeah right

What? Where did I make that comparison?

 

[Jaydip]

And we hardly know anything about it's performance,



What? Where did I make that comparison?

What kind of performance you are looking for?It is not meant for games.
You compared the die size dude.

 

[Red Hawk]

What kind of performance you are looking for?It is not meant for games.
You compared the die size dude.

He did compare the die size, but you bringing up manufacturing process was irrelevant. His point was that Fermi had a significantly larger die in proportion to Cypress/Evergreen than Tahiti/GCN has in proportion to GK104/Kepler in order to get better performance on the same manufacturing processes. That can be seen as either a con for Fermi or a plus for GCN.

 

[Jaydip]

He did compare the die size, but you bringing up manufacturing process was irrelevant. His point was that Fermi had a significantly larger die in proportion to Cypress/Evergreen than Tahiti/GCN has in proportion to GK104/Kepler in order to get better performance on the same manufacturing processes. That can be seen as either a con for Fermi or a plus for GCN.

It is relevant because if NV could manufacture Fermi on 28nm the size could be way smaller.

 

[HurleyBird]

What kind of performance you are looking for?It is not meant for games.

Do you have GPGPU benchmarks for GK110 then?

It is relevant because if NV could manufacture Fermi on 28nm the size could be way smaller.

You're confused. Fermi and Cypress are both 40nm parts.

 

[Jaydip]

Do you have GPGPU benchmarks for GK110 then?



You're confused. Fermi and Cypress are both 40nm parts.

Only theoretical values atm.You are missing the point, I am saying if Fermi could be manufactured using 28 nm it could be a lot smaller.Also Cypress was less focused on GPU compute compared to GCN.Fermi reserved lots of die spaces for GPU compute which was a waste for most of the gamers.I think GCN can't be really compared to GK104 atm, it should be really compared to GK110 if we get a Geforce derivative.

 

[Avalon]

Went with Kepler but I'm running a 7870

 

[ShadowOfMyself]

Until NV is the clear price/performance winner why should they be more favored? Many of us are swayed by performance and price or a ratio thereof. Please enlighten us in which area NV should be ahead of AMD where they aren't or haven't been in the polls?

This... That post could only have been made by an Nv fan, because anyone who has been here for the last 10 years, knows full well the community follows whatever is best at the time

Back in 2003, you might have said the same because there was nothing but joke threads about the leafblower aka Geforce FX

Jump to 2006, suddenly the whole forum is praising Nvidia for their amazing 8800 GTX

And so on

Right now, the 2 companies are pretty even, but as parvadomus posted above, and looking at factual evidence, GCN is the superior architecture, considering current cards

It might change with 8000/700 series, but right now, it is only logical that AMD is winning the poll

 

[3DVagabond]

Fermi vs Evergreen parrallels anyone?

Yeah, except Fermi used >100W more power and wasn't cheaper. Other than that, very similar.

 

[boxleitnerb]

Right now, the 2 companies are pretty even, but as parvadomus posted above, and looking at factual evidence, GCN is the superior architecture, considering current cards

It might change with 8000/700 series, but right now, it is only logical that AMD is winning the poll

What factual evidence? Price/performance is irrelevant when looking at architectures.
Problem is, the OP didn't specify what metrics we should look at.

When we look at compute, it massively depends what application is run. Some are pretty good on GK10x Kepler, some are abysmal. GK110 is not evaluated, too.

So I don't think there is an easy "Kepler or GCN" answer here that the OP would have liked.

 

[Red Hawk]

It is relevant because if NV could manufacture Fermi on 28nm the size could be way smaller.

So? Evergreen would be smaller on 28 nm too.

Only theoretical values atm.You are missing the point, I am saying if Fermi could be manufactured using 28 nm it could be a lot smaller.Also Cypress was less focused on GPU compute compared to GCN.Fermi reserved lots of die spaces for GPU compute which was a waste for most of the gamers.I think GCN can't be really compared to GK104 atm, it should be really compared to GK110 if we get a Geforce derivative.

Um...yeah, exactly. GCN is more compute-focused than Kepler, similar to how Fermi was more compute focused than Evergreen. The point is that GCN was able to accomplish this compute advantage (and general performance advantage) over Kepler while having a smaller die size ratiothan Fermi had over Evergreen. And if it's not fair to compare GCN (Tahiti) to GK104, was it never fair to compare GF100/110 to Cypress or Cayman?

 

[Jaydip]

So? Evergreen would be smaller on 28 nm too.



Um...yeah, exactly. GCN is more compute-focused than Kepler, similar to how Fermi was more compute focused than Evergreen. The point is that GCN was able to accomplish this compute advantage (and general performance advantage) over Kepler while having a smaller proportional die size than Fermi had over Evergreen. If it's not fair to compare GCN (Tahiti) to GK104, was it never fair to compare GF100/110 to Cypress or Cayman?

GCN is more compute focused yet it loses to Quadro 6000 , a 2 year old card.What benchmarks are you really referring to? compute performance on consumers cards is irrelevant which NV learned pretty quickly.I won't be surprised if AMD follows a similar suit.

 

[Red Hawk]

GCN is more compute focused yet it loses to Quadro 6000 , a 2 year old card.What benchmarks are you really referring to? compute performance on consumers cards is irrelevant which NV learned pretty quickly.I won't be surprised if AMD follows a similar suit.

Ok you're just trying to derail the discussion now. Anyways, the point is that GCN accomplishes a performance advantage over Kepler cards, compute or general performance, with less of a die size to performance ratio than Fermi took to get its performance lead over Evergreen. GCN can be applauded for that.

 

[Jaydip]

Ok you're just trying to derail the discussion now. Anyways, the point is that GCN accomplishes a performance advantage over Kepler cards, compute or general performance, with less of a die size to performance ratio than Fermi took to get its performance lead over Evergreen. GCN can be applauded for that.

No Dude, check my earlier post I didn't vote for either.My points are

1. GCN can't be compared to Kepler(Unless you want to discuss GK110) the same way Evergreen can't be compared to Fermi.

2.Both Evergreen and GK104 are gaming focused chips, they are not geared towards GPU compute.

3.Compute is irrelevant for consumer cards as NV learned it the hard way.They wasted a lot of die space for nothing.

4.When you use the term compute check the reviews of the professional cards.A two year old Fermi based Quadro 6000 is still ahead of Firepro 9000.Having great raw hardware means squat if the accompanying software is not up to the mark.

5. GCN architecture will be tested once it goes toe to toe with GK110.But GCN is a commendable architecture and I concur with that.

 

[Red Hawk]

No Dude, check my earlier post I didn't vote for either.My points are

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

1. GCN can't be compared to Kepler(Unless you want to discuss GK110) the same way Evergreen can't be compared to Fermi.

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.

2.Both Evergreen and GK104 are gaming focused chips, they are not geared towards GPU compute.

Ok.

3.Compute is irrelevant for consumer cards as NV learned it the hard way.They wasted a lot of die space for nothing.

No comment.

4.When you use the term compute check the reviews of the professional cards.A two year old Fermi based Quadro 6000 is still ahead of Firepro 9000.Having great raw hardware means squat if the accompanying software is not up to the mark.

And software has what to do with the architecture?

5. GCN architecture will be tested once it goes toe to toe with GK110.But GCN is a commendable architecture and I concur with that.

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