CUDA 5 & New Kepler Teslas

[GUIGuy]

http://www.theverge.com/2012/5/15/30223 ... k104-gk110

No word on pricing yet

[mbetke]

1500-2000 USD for the small Tesla

[gabrielefx]

more infos here:

http://www.vizworld.com/2012/05/nvidia- ... more-23799

http://pressroom.nvidia.com/easyir/cust ... xhtml=true

[Jaberwocky]

Looking at the info available.It sort of explains the rubbish performance of the kepler 680's up till now.Looks like your going to need Cuda 5 to get the most out of it.

quote:

"Dynamic Parallelism -- This capability enables GPU threads to dynamically spawn new threads, allowing the GPU to adapt dynamically to the data. It greatly simplifies parallel programming, enabling GPU acceleration of a broader set of popular algorithms, such as adaptive mesh refinement, fast multipole methods and multigrid methods.
Hyper-Q -- This enables multiple CPU cores to simultaneously use the CUDA architecture cores on a single Kepler GPU. This dramatically increases GPU utilization, slashing CPU idle times and advancing programmability. Hyper-Q is ideal for cluster applications that use MPI."

It also looks like Octane's going to have to be optimised to use multicores to get the most out of Kepler.

[glimpse]

Great presentation yesterday!

Well..few observations from what i've heard and read..so far:
"Windows 7 Support (Tesla M2070Q Only)" - Source

So, what does that say? NVidia is leaving small users on Fermi, those k10 and upcoming k20 models are left for enterprice =) i believe the demand for them is going very high & untill we see those in shops..it has to pass a bi of time..

About the speed =) Kepler is not bad at all. Peak on Single Precision of m2070q (Fermi based) is ~1Tflop, while Kepler has ~4,5tFlops on k10 model (that made from two gk104s).

[gabrielefx]

now it's more clear:

http://www.nvidia.com/object/tesla-servers.html

do we need double precision?

[glimpse]

now it's more clear:

http://www.nvidia.com/object/tesla-servers.html

do we need double precision?

No, we don't as so far as i know it's based on sp.

But as i mentioned on a post before, We probably will not be able to get those high performers just yet..=) it's written black on white, that there are no drivers for them..-and the targed application is different =)

[ChrisVis]

Hi guys,

I am new to Octane and the forum, this is my first post.
But I am very excited about the development in the Nvdia GPU market, as it just happens right now.

As I understand, the new Kepler GK104 Chip needs CUDA 5 for its best performance.
CUDA 5 preview version is available for developers since yesterday, the final Version will be out in Q3/2012.

Here is the press anounce, sorry only in german, as I am located in germany.
http://www.heise.de/newsticker/meldung/ ... 76477.html

About Single and double precision and the Tesla K10:

Tesla K10 uses the same GK104 chip (2 of them), then the GTX690 uses (yeah, its out since some days)!
And both Cards seem to have the same single presision performance!

Advantages over the GTX690 seems only to be the Memory (K10: 8GB, GTX690:4GB) and lower heat generation in terms of lower clock speeds (K10: 745Mhz per GPU - 2,28 TFlops per GPU, Geforce 680: 915Mhz, whats also a performance disadvantage for the K10!). Both cards have 3072 CUDA Cores!

Nvidia markets the K10 only as single precision card, not as double precision card, cause it`s double precision performance is very weak (190Glps)!!
Can be read here, only german, too, sorry.
http://www.heise.de/newsticker/meldung/ ... 75385.html

Tesla K20 with high double precision performance and GK110 Keppler chip will be out in End of 2012, and will have 2880 CUDA Cores on a single chip!! The first Geforce GTX with GK110 will be out in early 2013.
http://www.heise.de/newsticker/meldung/ ... 76464.html

So what does it mean for octane users!?

We dont need a Tesla K10 for 2000 bucks, instead a Geforce GTX690 would have an even better performance for about 999 bucks!

And: We may have to wait 3 or 4 month from now for CUDA 5 release and a new octane release, that ist fully supporting CUDA 5. Then I think, that the 3072 CUDA Cores from the Kepler will at least behave as the Fermi Cores and we will see a huge performance step up, like 4x to 6x from what we are used to with 1xGTX580.

What do you think?

[glimpse]

Hi guys,

I am new to Octane and the forum, this is my first post.
But I am very excited about the development in the Nvdia GPU market, as it just happens right now.

...

We dont need a Tesla K10 for 2000 bucks, instead a Geforce GTX690 would have an even better performance for about 999 bucks!

And: We may have to wait 3 or 4 month from now for CUDA 5 release and a new octane release, that ist fully supporting CUDA 5. Then I think, that the 3072 CUDA Cores from the Kepler will at least behave as the Fermi Cores and we will see a huge performance step up, like 4x to 6x from what we are used to with 1xGTX580.

What do you think?

Firstly, hi!

Well, 580 has ~849 GFlops (single precision floating point),590 - asume about 1500 (?), new kepler K10 4577GFlops (2288 GFlops per GPU).

So, at best..if 690 will be as good as K10..yes roughly 6x over 580, but only if..Developers will be able to get this power of that hardware..

Let's hope for the best =) if with 690 we'll be able to outperform 4x580 rig - this is going to be epic!

[ChrisVis]

Hi glimpse!

Yeah, if one gtx690 could outperform a 4xgtx580 rig... then just think of a 4xgtx690 rig *g*
But i dismissed a point: the 4GB of the gtx690 might has to be splitted to 2x 2GB for each GPU in theory...
so this would not be better in terms of RAM usage with octane than a GTX580 with 3GB of RAM, right? (Because every single GPU only is capable of using 2GB RAM... or is the memory somehow shared for both GPUs on the card?)
Maybe they will come out with an 8GB Version of the gtx690 later?

Another question: I am sure its somewhere else on the Forum, and i just didn`t find it yet, but is there a real good and affordable solution for an external GPU rig out there anywhere? Besides onestopsystems and cubix?
I am searching for a DIY kit and only need a good and working PCI-E Expansion backplane card system.

Found this one, but seems to be price intensive:
http://cyclone.com/products/expansion_b ... /index.php

And this one, that looks very affordable (about 600$), but does it have the performance needed?
http://www.adnaco.com/products/s1b/?gcl ... zQodWktcDw

It says: "Any type of PCI Express cards can be used including audio, video, graphics, USB, FireWire, SATA, data-acquisition, network and others." And it is wired through fibre cable... so the rig system could be 10-100 Meter away.

I`d like to built my own GPU-Rig with watercooling to keep it quiet and cool, because I don`t have the option to place it in another room.
Maybe I start with a 4xGTX580 3GB watercooled system now and add another Rig with GTX680 or GTX690 later?
I need more performance right now in the next 2 month to do a job, so it wouldn`t be wise to buy a 4xGTX680 rig at the moment, right?

Not an easy decision with all the news at the moment...