Nvidia Kepler - Page 3

nucleon · 2012-03-22, 13:07

Reviews are out:

http://www.anandtech.com/show/5699/n...gtx-680-review

Compute benchmarks are a little disappointing.

-- Craig

flashjh · 2012-03-22, 13:22

Well, newegg posted them...

http://www.newegg.com/Product/Produc...tx+680&x=0&y=0

aaronhaviland · 2012-03-22, 14:09

Quote:

Originally Posted by nucleon

Reviews are out:

http://www.anandtech.com/show/5699/n...gtx-680-review

Compute benchmarks are a little disappointing.

-- Craig

Indeed:

http://www.brightsideofnews.com/news....aspx?pageid=4

KyleAskine · 2012-03-22, 14:53

It sounds like nVidia's plan is to make sure that you cannot buy the gamer boards anymore and use them for compute projects. They are going to try to push compute users to much pricier Tesla and Quadro cards.

Dubslow · 2012-03-22, 15:14

I'm wondering how the GK110 will perform compute-wise. In the meantime, does anybody feel like shelling out to try one?

Prime95 · 2012-03-22, 17:21

Quote:

Originally Posted by KyleAskine

They are going to try to push compute users to much pricier Tesla and Quadro cards.

or AMD Radeon cards

flashjh · 2012-03-22, 18:43

Quote:

Originally Posted by Dubslow

I'm wondering how the GK110 will perform compute-wise. In the meantime, does anybody feel like shelling out to try one?

Got my 680 ordered. Hopefully will ship today and get here Monday. Based on the reviews I've seen here, I'm a bit concerned it won't perform much better than a 580 on CUDA. If not, I'm sure I'll be able to sell it and get a GK110 later.

rcv · 2012-03-22, 22:05

I haven't read the actual specs, but I did read those reviews. Assuming I interpret the reviews correctly, I'm disappointed. When a card runs at a much slower clock speed (than a 580) and it has half the number of multiprocessors (as a 580), and it doubles "some" of the circuitry, it's hard to see how it's compute performance can be competitive.

Of particular concern to me is the fact that they did not double the fast shared memory / L1 cache. You still get 64K of this resource per multiprocessor. The 580 came with 16 multiprocessors. The 680 has 8 multiprocessors. I think that means half the shared memory and half the L1 cache.

My sieving algorithms (mentioned elsewhere) scream with enough shared memory. Where they top out is proportional to the amount of shared memory. [The 680 could perform like an overclocked 560Ti.]

mfaktc, as I recall, doesn't make much use of shared memory or double-precision, so I suspect it will run well on the 680.

NVIDIA's manuals urge programmers to use special techniques to do matrix operations most efficiently using shared memory. The better you heeded their advice, the better you will be punished for it. [Just sayin'.]

I'm hoping I missed something that makes these concerns all wrong!!

Dubslow · 2012-03-22, 22:46

Keep in mind that the 680 is a GK104, so many senses it is more closely related to the 560 Ti than the 580; their branding is slightly confusing. (Also, it has a much higher core clock, 1000 vs. 7xx, and thought it had more CUDA cores as well...)

pinhodecarlos · 2012-03-22, 22:59

Just by looking at cards TDP I don't understand your worries. In my point of view, energy efficiency one, making at least the same with less energy is good.

bcp19 · 2012-03-23, 00:23

Quote:

Originally Posted by KyleAskine

It sounds like nVidia's plan is to make sure that you cannot buy the gamer boards anymore and use them for compute projects. They are going to try to push compute users to much pricier Tesla and Quadro cards.

The price alone make the Quadro and Tesla cards unpalatable to many. They also seem to be geared towards professional use in things like CAD/CAM and Video editing. Performance wise, the Quadro 5000 is comparable to a GTX 560, but at about 6x the cost. The question foremost in mind is what benefit is there for this added cost? While it's 2.5G vs 1G, memory seems to have little/no benefit for mfaktc, so would it have an impact on CUDALucas? It'd be great to have benchmarks for mfaktc and CUDALucas on both Quadro and Tesla to see what thier strengths are.
That said, I'm eagerly awaiting flash's report on the performance of the 680 in (hopefully) both mfaktc and CUDALucas.

2012-03-22, 13:22	#24
flashjh "Jerry" Nov 2011 Vancouver, WA 1,123 Posts	Well, newegg posted them... http://www.newegg.com/Product/Produc...tx+680&x=0&y=0 Last fiddled with by flashjh on 2012-03-22 at 13:27

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Nvidia GTX 745 4GB ???	petrw1	GPU Computing	3	2016-08-02 15:23
Nvidia Pascal, a third of DP	firejuggler	GPU Computing	12	2016-02-23 06:55
Pitfall when upgrading to Windows 10 with Fermi vs Kepler/Maxwell	Brain	Hardware	16	2015-11-26 10:24
AMD + Nvidia	TheMawn	GPU Computing	7	2013-07-01 14:08
What can I do with my nvidia GPU?	Surge	Software	4	2010-09-29 11:36

2012-03-22, 13:07	#23
nucleon Mar 2003 Melbourne 5×103 Posts	Reviews are out: http://www.anandtech.com/show/5699/n...gtx-680-review Compute benchmarks are a little disappointing. -- Craig

2012-03-22, 14:53	#26
KyleAskine Oct 2011 Maryland 2·5·29 Posts	It sounds like nVidia's plan is to make sure that you cannot buy the gamer boards anymore and use them for compute projects. They are going to try to push compute users to much pricier Tesla and Quadro cards.

2012-03-22, 15:14	#27
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3·29·83 Posts	I'm wondering how the GK110 will perform compute-wise. In the meantime, does anybody feel like shelling out to try one?

2012-03-22, 22:05	#30
rcv Dec 2011 227₈ Posts	I haven't read the actual specs, but I did read those reviews. Assuming I interpret the reviews correctly, I'm disappointed. When a card runs at a much slower clock speed (than a 580) and it has half the number of multiprocessors (as a 580), and it doubles "some" of the circuitry, it's hard to see how it's compute performance can be competitive. Of particular concern to me is the fact that they did not double the fast shared memory / L1 cache. You still get 64K of this resource per multiprocessor. The 580 came with 16 multiprocessors. The 680 has 8 multiprocessors. I think that means half the shared memory and half the L1 cache. My sieving algorithms (mentioned elsewhere) scream with enough shared memory. Where they top out is proportional to the amount of shared memory. [The 680 could perform like an overclocked 560Ti.] mfaktc, as I recall, doesn't make much use of shared memory or double-precision, so I suspect it will run well on the 680. NVIDIA's manuals urge programmers to use special techniques to do matrix operations most efficiently using shared memory. The better you heeded their advice, the better you will be punished for it. [Just sayin'.] I'm hoping I missed something that makes these concerns all wrong!!

2012-03-22, 22:46	#31
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3×29×83 Posts	Keep in mind that the 680 is a GK104, so many senses it is more closely related to the 560 Ti than the 580; their branding is slightly confusing. (Also, it has a much higher core clock, 1000 vs. 7xx, and thought it had more CUDA cores as well...)

2012-03-22, 22:59	#32
pinhodecarlos "Carlos Pinho" Oct 2011 Milton Keynes, UK 2×5×11×47 Posts	Just by looking at cards TDP I don't understand your worries. In my point of view, energy efficiency one, making at least the same with less energy is good.