Nvidia to launch Tesla Personal Supercomputer

[Hrvoje Simic]

4 TFLOPS for $10k.

link

 

[Martin Weiss]

It even says "Available now for under 10.000 USD".

960 Cores!

Could this machine be used for R3D files today - or do we need extra hard/software before that becomes a reality?

28k seems also totally possible

 

[Pawel Achtel]

It even says "Available now for under 10.000 USD".

960 Cores!

Could this machine be used for R3D files today - or do we need extra hard/software before that becomes a reality?

28k seems also totally possible

It looks you need to write custom code to take advantage of those boards, but you're right regarding 28k. The architecture should allow scaling speed almost linear for graphics applications as portions of images could be processed easily in parallel. So, it would be literally almost 250 times faster than desktop. Also, it should not be too diffificult to get the code ported across.

There is still a problem: how do you display the damn thing! I can't even display 4k

 

[Dave Blackham]

What about Apple OS X support, its non Apple hardware but if its possible on this then surly that would be over run with orders.

Dave

 

[Jeff Kilgroe]

OSX or not, this thing is going to sell like hotcakes... If the developer tools are robust enough and the learning curve isn't too steep.

This is why RED needs to get CUDA support into the R3D SDK, ASAP.

 

[JanneJansson]

If linux kernel recognize all core, then you just need REDline for linux to get goin.
..but probably you need to go in to the Nvidia CUDA API and build your own app.

But a linux REDline version would be a good start!
Cheers!

 

[michael zaletel]

Isn't storage I/O more important than processing speed for 28k?

Curious...

-shooter

 

[Adam Glick]

Getting the necessary storage bandwidth will be the easy part.

As always, processing will be the challenge.

 

[Thomas Mathai]

Isn't storage I/O more important than processing speed for 28k?

Curious...

-shooter

Maybe use infiniband. Wonder if solid state storage can sustain the speeds

 

[Gavin Greenwalt]

28k was only 225MBs

I could build a box which does that for 2 grand easy.

 

[will_w]

Tesla is FPGA

Tesla is FPGA

The Tesla system is an FPGA. Basically this means they've allocated physical space on the chip that would usually be used for decision making and branching to parallel processes. In other words - there are some times when you want one chip that does 100 things. This is 100 chips that do 1 thing. There is NO GRAPHICS associated with the Tesla. It is strictly an FPGA. It is often used to accelerate a graphics process because drawing to a screen inherently means repeating the same thing over and over - which is what this kind of board is good at.

Re: Mac compatibility
It uses the CUDA language natively which means it is platform independant. CUDA is a list of addons to the C language.

If you want a graphics card that uses the Tesla chips - look at the nVidia Quadra series.

 

[Mat@imageWork]

Yes CUDA is a little bit young, but nvidia makes it easier to compile in C existing application. Off course you will need to have access to the complete code source and re compile it with CUDA in order to get performance improvement. Go on youtube and tape CUDA, there is a forum there, video and demo. Some performance are x10 and even more. For rendering r3d, compositing 4K, that's the way to go. Boxxlab, I know you are supporting tesla supercomputer, could you share some test and result?

Many thanks

 

[Adam Glick]

Yes, we just showed our implementation of the PSC at the Nvidia booth here in Austin at SC08 - the big supercomputing conference.

The PSC is of immediate interest to researchers, engineers and scientists who need massive floating point performance and would otherwise need to share time on a "traditional" departmental supercomputer.

As far as using FPGA logic arrays (such as the PSC and GPUs) for decompressing/debayering "RAW" camera data streams -it's obviously a viable approach as we all have seen a number of software companies basing their products on that.

Apparently there are also trade-offs doing it this way that make the alternate method (software decode on a multi-core CPU) equally attractive or beneficial.

 

[Mat@imageWork]

Yes, we just showed our implementation of the PSC at the Nvidia booth here in Austin at SC08 - the big supercomputing conference.

The PSC is of immediate interest to researchers, engineers and scientists who need massive floating point performance and would otherwise need to share time on a "traditional" departmental supercomputer.

As far as using FPGA logic arrays (such as the PSC and GPUs) for decompressing/debayering "RAW" camera data streams -it's obviously a viable approach as we all have seen a number of software companies basing their products on that.

Apparently there are also trade-offs doing it this way that make the alternate method (software decode on a multi-core CPU) equally attractive or beneficial.

Thanks. Everything is matter of compiling the source code to accomodate render farmed workflow (traditional) or parallel computing (PSC). I do believe PSC could take a clear avantage as long as R3D is based on JPEG2000 compression algorythm and parallel computing could habe a hudge impact in processing time. And as this time, nether one or the other solution have yet been done, with make the use of RED Cam in my case (4K stereoscopic film) a "pain in the ..."

 

[harjotsingh]

Wish i could use all that hardware but for now only Linux is the only os that can handle that much power. 28k is not possible because of software not hardware. Think about rendering a 28k canvas in aftereffects. Right now aftereffects only allows 4 processors. so once that is done. where are you going to put the 28k footage. love to dream of a day where 28k will be like 1080

 

[Andrae Palmer]

Yes, we just showed our implementation of the PSC at the Nvidia booth here in Austin at SC08 - the big supercomputing conference.

The PSC is of immediate interest to researchers, engineers and scientists who need massive floating point performance and would otherwise need to share time on a "traditional" departmental supercomputer.

As far as using FPGA logic arrays (such as the PSC and GPUs) for decompressing/debayering "RAW" camera data streams -it's obviously a viable approach as we all have seen a number of software companies basing their products on that.

Apparently there are also trade-offs doing it this way that make the alternate method (software decode on a multi-core CPU) equally attractive or beneficial.

Can you enlighten us on what some of the trade offs are? I am new to all of this and would like to know if an NVIDIA Tesla or Quadra Plex would solve all of our high resolution debayering problems.

 

[Harva Raj]

if a windows vista/xp application is multi processing capable does it take advantage of the 960 cores? or do you have to write a special code for it? does it have a problem running vista/xp?

 

[Andrae Palmer]

if a windows vista/xp application is multi processing capable does it take advantage of the 960 cores?

No

do you have to write a special code for it?

CUDA

does it have a problem running vista/xp?

Supported Platforms

Microsoft® Windows® XP 64-bit and 32-bit (64-bit recommended)
Linux® 64-bit and 32-bit (64-bit recommended)
Red Hat Enterprise Linux 4 and 5
SUSE 10.1, 10.2 and 10.3

http://www.nvidia.com/object/tesla_supercomputer_tech_specs.html

 

[laguun]

Yes, we just showed our implementation of the PSC at the Nvidia booth here in Austin at SC08 - the big supercomputing conference.

Boxxlabs is exhibiting at supercomputing conferences?
i thought boxx is a standard PC integrator and is selling assembled regular TYAN motherboards?

did i miss something?

p.s. is there an ETA when you will sell dual nehalem boxes and did you decide on the motherboard?

 

[Mat@imageWork]

The red sdk doesn't allow us to access the debayering process is coded, so no chance actually to recomplie this for CUDA render out the r3d...