Geforce GTX Titan 6GB - Page 5

LaurV · 2013-02-25, 09:20

There is no "780".

This was how the chip was named by the gossipers, when it was rumored long before it appeared. Nvidia have Fermi and Kepler architectures, for which they made "consumer" chips, GeForce (respective "GTX 5xx" for Fermi and "GTX 6xx" for Kepler) and "professional" chips (Teslas, "2xxx" for Fermi and "Kxx" for Kepler respectively).

The "7xx" series would correspond to (next-to-come) Maxwell architecture, which is not released yet, barely rumored.

IMHO, Titans are just "non-conform" K20X, if you look to the characteristics, they are quite similar. If you look here their performance is not much over a top Fermi or Kepler "top" card (I don't talk about dual chips like 690 or 590) even for gaming! therefore it would make totally no sense to be "the launching of the 7xx series, top down" like this guy says in this article, he has no idea what he is gibbering about. A "top Maxwell" would jump much higher. I am still waiting for it.

This "non-conform" story does not have to scary anybody. Teslas are MONSTERs of hardware, with few billions - BILLIONS - "building blocks", so with even a fail rate of 0.0001 ppm (parts per million) - which is common for this industry, you have (about) one chip in 10 which is not totally working, but those are STILL perfectly usable, if you "cut out" this or that block, or if you use underclock this or that part. This is what a "GTX560 Ti 448 cores" it is, just a "rejected" 570, and I don't know anybody to complain about 560Ti's, contrarily, everybody seems to like it, as its performance is higher then a normal 560, for a price lower then a 570.

That is what Titan is, to have an analogy, IMHO.

ixfd64 · 2013-02-25, 23:13

FYI: Hexus is giving one away: http://hexus.net/tech/features/graph...orce-gtx-titan

ET_ · 2013-02-26, 08:19

Quote:

Originally Posted by ixfd64

FYI: Hexus is giving one away: http://hexus.net/tech/features/graph...orce-gtx-titan

Welcome to the team...

Luigi

ixfd64 · 2013-02-26, 17:17

G.Skill is also giving one away: http://www.tweaktown.com/news/28770/...an-/index.html

kladner · 2013-02-26, 21:02

Quote:

Originally Posted by ixfd64

G.Skill is also giving one away: http://www.tweaktown.com/news/28770/...an-/index.html

Grrr! Must sign up on Facebook.

henryzz · 2013-02-26, 22:40

Quote:

Originally Posted by axn

It doesn't necessarily have to be a contradiction. Titan has peak thruput of 1.5Tflops vs 1.2 for K20. Also mem bandwidth is 288 GB/s vs 208 for K20. While the benchmark gives 2.5x performance in FFT, CuLu would also spend a significant chunk of the time in the carry-propagation routine(s) which is entirely memory bottlenecked [if my experience with GeneferCUDA timing is anything to go by].

Also, IIRC, the benchmark was using cuFFT 5.0 & CuLu uses cuFFT 3.2. Maybe that also causes suboptimal performance?

I had estimated the actual performance in CuLu to be about 1.75x-2x rather than the 2.5x given for raw FFT performance.

Considering all this, I wouldn't be surprised if K20 only gave a 10% performance boost over 580 (which would be reported as "about the same").

/SWAG

In this post axn mentions that CudaLucas uses cuFFT 3.2. Is there a particular reason we haven't upgraded to the latest version?

Dubslow · 2013-02-26, 22:56

Quote:

Originally Posted by henryzz

In this post axn mentions that CudaLucas uses cuFFT 3.2. Is there a particular reason we haven't upgraded to the latest version?

As far as I can tell from having compiled it a bazillion times, it uses whatever version your CUDA SDK provides -- which is presumably the same version as the SDK. So using the 3.2 SDK would get you cuFFT 3.2, while using the 4.1 SDK would get you the 4.1 cuFFT library. (Now, whether or not there are actually and changes between those "versions", who knows.)

henryzz · 2013-02-26, 23:42

Quote:

Originally Posted by Dubslow

As far as I can tell from having compiled it a bazillion times, it uses whatever version your CUDA SDK provides -- which is presumably the same version as the SDK. So using the 3.2 SDK would get you cuFFT 3.2, while using the 4.1 SDK would get you the 4.1 cuFFT library. (Now, whether or not there are actually and changes between those "versions", who knows.)

The cuFFT page on nvidia's website in a graph showed improvement between 4.0 and 4.1 especially for slightly unusual fft lengths(including things like 5s)

ET_ · 2013-02-27, 07:50

Quote:

Originally Posted by henryzz

The cuFFT page on nvidia's website in a graph showed improvement between 4.0 and 4.1 especially for slightly unusual fft lengths(including things like 5s)

Time for some benchmarks?

Luigi

henryzz · 2013-02-27, 12:32

Quote:

Originally Posted by ET_

Time for some benchmarks?

Luigi

It would be nice if someone who could compile cudalucas could do some with different versions.

Redarm · 2013-02-27, 22:48

got one who tried mfactc

http://www.abload.de/img/ertertrelffg2.jpg

driver is not clucas-compatible for now

2013-02-25, 09:20	#45
LaurV Romulan Interpreter "name field" Jun 2011 Thailand 41·251 Posts	There is no "780". This was how the chip was named by the gossipers, when it was rumored long before it appeared. Nvidia have Fermi and Kepler architectures, for which they made "consumer" chips, GeForce (respective "GTX 5xx" for Fermi and "GTX 6xx" for Kepler) and "professional" chips (Teslas, "2xxx" for Fermi and "Kxx" for Kepler respectively). The "7xx" series would correspond to (next-to-come) Maxwell architecture, which is not released yet, barely rumored. IMHO, Titans are just "non-conform" K20X, if you look to the characteristics, they are quite similar. If you look here their performance is not much over a top Fermi or Kepler "top" card (I don't talk about dual chips like 690 or 590) even for gaming! therefore it would make totally no sense to be "the launching of the 7xx series, top down" like this guy says in this article, he has no idea what he is gibbering about. A "top Maxwell" would jump much higher. I am still waiting for it. This "non-conform" story does not have to scary anybody. Teslas are MONSTERs of hardware, with few billions - BILLIONS - "building blocks", so with even a fail rate of 0.0001 ppm (parts per million) - which is common for this industry, you have (about) one chip in 10 which is not totally working, but those are STILL perfectly usable, if you "cut out" this or that block, or if you use underclock this or that part. This is what a "GTX560 Ti 448 cores" it is, just a "rejected" 570, and I don't know anybody to complain about 560Ti's, contrarily, everybody seems to like it, as its performance is higher then a normal 560, for a price lower then a 570. That is what Titan is, to have an analogy, IMHO. Last fiddled with by LaurV on 2013-02-25 at 09:25

2013-02-27, 22:48	#55
Redarm Apr 2012 Berlin Germany 3·17 Posts	got one who tried mfactc http://www.abload.de/img/ertertrelffg2.jpg driver is not clucas-compatible for now Last fiddled with by Redarm on 2013-02-27 at 22:51

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Titan's Best Choice	Brain	GPU Computing	30	2019-10-19 19:19
Titan Black	ATH	Hardware	15	2017-05-27 22:38
Is any GTX 750 the GeForce GTX 750 Ti owner here?	pepi37	Hardware	12	2016-07-17 22:35
Nvidia announces Titan X	ixfd64	GPU Computing	20	2015-04-28 00:27
2x AMD 7990 or 2x Nvidia Titan ??	Manpowre	GPU Computing	27	2013-05-12 10:00

2013-02-25, 23:13	#46
ixfd64 Bemusing Prompter "Danny" Dec 2002 California 2504₁₀ Posts	FYI: Hexus is giving one away: http://hexus.net/tech/features/graph...orce-gtx-titan

2013-02-26, 17:17	#48
ixfd64 Bemusing Prompter "Danny" Dec 2002 California 2³×313 Posts	G.Skill is also giving one away: http://www.tweaktown.com/news/28770/...an-/index.html