Some interesting Kepler leaks

[xsistor]

A few days old but hey... Quite interesting.

http://lenzfire.com/2011/12/nvidia-...l-focus-mainly-on-performance-per-watt-81669/

Nvidia claimed that, Fermi architecture would achieve typical double precision (DP) performance of 1.5GFLOPS per watt. Also in Kepler architecture the performance per watt is increased to about 3 or 4 times, hence, we can expect 1.125TFLOPS – 1.50TFLOPS of DP performance by a chip with 250W thermal design power.

Echelon will be its future chip with 1024 cores and 10 teraflops power which will mainly focus on high performance computing. Also it will run on 128 mini-Streaming Multiprocessors (SM) each consisting of 8 cores

 

[chaotic_uk]

but will it play crysis ?

 

[Sin_Chase]

Boring computing stuff.

Really not liking the way focus is being applied to HPC, while I understand it's a huge money spinner it's not as exciting for us in the gaming world.

 

[drunkenmaster]

Rubbish info is rubbish, the 3-4DP/W increase does in no way suggest how much DP they are planning on providing. IE if a 580gtx can do 500dp flops for 300W, the new card they are comparing it to might do 250dp flops, but at 50W that would be 3x dp flops/w better, but no actual improvement in performance.

the 3-4 increase is also going to be absolutely BS, it's Nvidia and a slide about future performance, they've been promising stuff they haven't delivered in these things for 5+ years.

You should get 2x DP performance per watt from a bog standard die shrink anyway, without trying, shrink an identical 512sp 580gtx to 28nm with perfect scaling and you'd expect a little over 2x dp/w without a single change to architecture. A slight architectural change could potentially raise that to 2.5-3 very easily, but its Nvidia, and PR and promises and I'd be surprised if we saw anything over 2.5x in reality, and unsurprised if we see less than a 2x increase.

 

[SFFNoob]

Someone get some more info! Be it on sheer performance!

 

[Freddie1980]

Good news if your into High performance computing I guess.

 

[Laher]

just like 7970, achieved their goals in computing....provided no real gaming improvement worthy of my ££. Roll on 2013

 

[Rroff]

Nothing actually new information wise on that page other than the Echelon stuff which has no real relevance to Kepler and was part of a presentation nVidia did recently on theoretical super computer applications.

 

[Duff-Man]

You should get 2x DP performance per watt from a bog standard die shrink anyway, without trying

Not really. Assuming the same architecture, there is nothing to suggest DP performance would scale any differently than SP performance.

The 7970, for example, has the 947 GFLOPS of DP performance, compared to 683 GFLOPS for the 6970 (an increase of 38%). This also comes with a small increase in power draw (5-10%), so a per-Watt efficiency change of +(30-35)%. There is nothing to suggest DP efficiency has changed out of proportion to SP.

So, from a pretty damn big process shrink (40nm->28nm), AMD has managed an overall increase in per-Watt DP floating point performance of around 30-35%. Why is it that you think that Nvidia will be able to double the DP performance "without even trying" when AMD have struggled to get one third of that?

 

[Jokester]

It's DP/W that DM is talking about, but yeah your figures still show that a x2 increase didn't happen for DP/W either for AMD.

 

[Duff-Man]

It's DP/W that DM is talking about, but yeah your figures still show that a x2 increase didn't happen for DP/W either for AMD.

Both SP and DP floating point performance have increased by less than 40% (compared to 6970)

Power draw has risen by 5-10% under load (compared to 6970).

This implies less than 35% improvement in overall floating-point performance per watt. There is nothing to suggest that pure double-precision computations would stray away from this trend.

This is miles away from the suggestion that it is "easy" to "double" DP performance-per-Watt with a die shrink.


It may or may not be the case that Nvidia can show a factor of 2-3 increase in DP FP performance with Kepler. But suggesting that this can be achieved every time a die shrink occurs, "without trying" is pretty stupid. If that were the case, AMD would not have been so far off the mark. As with just about everything else related to GPU design, the reality is far more complex than "certain people" would like to believe.

 

[xsistor]

Both SP and DP floating point performance have increased by less than 40% (compared to 6970)

Power draw has risen by 5-10% under load (compared to 6970).

This implies less than 35% improvement in overall floating-point performance per watt. There is nothing to suggest that pure double-precision computations would stray away from this trend.

This is miles away from the suggestion that it is "easy" to "double" DP performance-per-Watt with a die shrink.


It may or may not be the case that Nvidia can show a factor of 2-3 increase in DP FP performance. But suggesting that this can be achieved every time a die shrink occurs, "without trying" is pretty stupid. If that were the case, AMD would not have been so far off the mark. As with just about everything else related to GPU design, the reality is far more complex than "certain people" would like to believe.

It's just DrunkenMaster talking out of his rear end as usual. Just ignore it, mate. Not worth the trouble

 

[Purgatory]

You should get 2x DP performance per watt from a bog standard die shrink anyway, without trying,

Drunkmaster we did this already with the AMD 7970 and the clear facts state that is wrong and even your 80% improvement didn't materialise.

I think we all have to just wait and see what the real improvements will be for gaming and compute when high end cards come out now.

The 7970 to me is an overclocked GTX 580 and nothing more really. DirectX 11.1 is not out till windows 8 and reality it adds nothing for the consumer and most of the updates are for the developers. PCI-E 3 humm another one that will probably only benefit people doing crossfire or until 7990 is out and we will see if it makes any real world difference. 3GB ram humm another thing not required yet if you game at 1920 x 1200 or 1080.

Lets see what Nvidia can do but i'm not raising my hopes, die shrinks and new architectures especially have to prove they make things better for us before we all claim it will be better because it has new architecture blaa blaa ... Bullbozer is a new architecture CPU from AMD ... Is it any better then previous architecture ? The only impressive thing I spotted on the 7970 is the Crossfire scaling, but again I don't do crossfire or SLI and I would never put my money in the hands of AMD regarding crossfire drivers and SLI also has its problems too from what I understand and this is why I always buy one top end single GPU card, I was even tempted to get a GTX 590 and then thought do I really want to have headaches in the future ? I thought naa get a GTX 580 for now and when a new single GPU highend card comes out again that shows good gains update again.


Merry Christmas all and hope you all had a fantastic day yesterday and enjoying your Boxing day.

 

[Rroff]

Sheesh what is it with people, DM and others including myself said - to paraphrase:

"Its usual with a generation leap (rather than a refresh) to see a performance increase of about 70% from previous generation top end card to next generation top end card"

Something that can be see to happen in most cases in the past. AFAIK he never claimed that it would actually be a ~70% performance increase - if he did I missed that.

 

[Purgatory]

Sheesh what is it with people, DM and others including myself said - to paraphrase:

"Its usual with a generation leap (rather than a refresh) to see a performance increase of about 70% from previous generation top end card to next generation top end card"

Something that can be see to happen in most cases in the past. AFAIK he never claimed that it would actually be a ~70% performance increase - if he did I missed that.

http://forums.overclockers.co.uk/showpost.php?p=20538014&postcount=12


And here how he changed his tune and went back claiming it is 80% faster then a 5870.. when reality was he was claiming before it will be 2x faster then a 6970.


http://forums.overclockers.co.uk/showpost.php?p=20853554&postcount=74



Anyway mate does not matter... Enjoy the holidays guys... But people were claiming 80% to 2x the performance of a 6970 at one point and stating they will be like £300-350 a card when they come out.. When I said they will be closer to £500 and a lot smaller increase in performance than they are expecting, people then screamed murder and said I have no idea what i'm talking about and same to others that realised this was not going to be the big 28nm card this year. You will see if you read and fully open them posts what was said and there was a few more if you look back mate.

 

[Rroff]

Oh OK I stand corrected, I only remember the post where he said 70% was traditionally what we'd see.

 

[SFFNoob]

I've missed something here, haven't I?

Though I'm interested on the PCI-E3 note, you say these GPU's haven't saturated the PCI-E2 bandwidth?

 

[Freakbro]

I've missed something here, haven't I?

Though I'm interested on the PCI-E3 note, you say these GPU's haven't saturated the PCI-E2 bandwidth?

I was interested in that too, so I did a little looking around and it seems the current cards can just about saturate a x8 PCIe 2.0 slot, so no, they will be nowhere near saturating a full x16 slot.

So its unlikely the next gen cards will come close either.

Anyone else confirm or refute this?

 

[Duff-Man]

Anyone else confirm or refute this?

Yeah, the tiny performance difference between 8x/8x and 16x/16x SLI or crossfire setups suggests that the bandwidth of PCI-e 8x is barely being saturated with current-gen cards.

I can't see there being much of a performance bump heading to PCI-e 3.0 with this generation. I'd hazard a guess at a maximum of maybe 1-2% maximum improvement over a PCI-e 2.0 16x slot, and even then only in certain circumstances.

 

[xsistor]

I was interested in that too, so I did a little looking around and it seems the current cards can just about saturate a x8 PCIe 2.0 slot, so no, they will be nowhere near saturating a full x16 slot.

So its unlikely the next gen cards will come close either.

Anyone else confirm or refute this?

For graphics, it doesn't look like it, though the double GPU cards may actually see benefits. The NVIDIA high end may also see benefit if they really do continue to make monolithic GPUs like the 580 with their 28nm process.

The additional bandwidth will be useful for compute purposes however.

Also, I remember reading some time ago (before PCI-SIG completed the specification) that PCI-E 3 will allow higher power draw from the socket, so that might also help with the super-high end and dual GPU cards.