NVIDIA CUDA dump from SC2007

Tim Murray · 13-Nov-2007, 04:44

Huge amount of stuff I'm still parsing, so you'll see some more news/real analysis soon--just wanted to get this out there first.

- CUDA 1.1 beta available for registered developers--debugging on the GPU, visual profile, async support, future Windows drivers will have CUDA support in the main branch, stream management? Lots and lots and lots of stuff.

- The eight hour CUDA presentation from SC07. Still going through it, but they're still saying DP next month. So far, I've read a lot of things that I've already seen, but give me some time.

And of course, thanks to the lovely person who pointed this out to me

Rufus · 13-Nov-2007, 04:59

And for those of us into a bit of theory wankery, there's a beta release of CUDPP (Cuda Data Parallel Primatives) up here: http://www.gpgpu.org/developer/cudpp/

It's a library consisting of scan, sort, and pack. Doesn't sound like much, but there is a huge class of parallel algorithms that depend on nothing but those 3, and if those 3 are done well the algorithms will Just Work.

Tim Murray · 13-Nov-2007, 05:50

I just read the optimization presentation. My head hurts a lot.

Lux_ · 13-Nov-2007, 08:20

Last link is broken (Wave Propagation Through Earth)?

Rufus · 13-Nov-2007, 18:18

A bit of bad news? http://www.hpcwire.com/hpc/1889153.html

Quote:

In early 2008, NVIDIA plans to release the next generation of Tesla offerings. These devices will include double precision floating point support, giving the product line a much greater range of applicability on scientific codes. Presumably the new GPUs will be based on the same 65nm process technology incorporated into the company's recently released G92 graphics chip.

DP pushed back to "early 2008" instead of this year?

cho · 14-Nov-2007, 03:21

there is a paper about DP by University of Virginia and NVIDIA Research :

http://www.cs.virginia.edu/~ktd3q/docs/ssr_gpu_ije.pdf

"We show that SSR can be used effectively in programmable graphics architectures to allow double-precision computation without affecting the performance of single precision calculations and to increase fragment shader performance with a minimal impact on chip area."

Tim Murray · 14-Nov-2007, 03:22

Well...

Quote:

Originally Posted by Massimiliano Fatica

G92 does not support double precision

So unless he's making a distinction between G92 or G92GL... hi G90?

Geo · 07-Jan-2008, 17:36

http://blogs.sun.com/HPC/entry/conso...idia_tesla_for

Is that part of the dump upstream or something new?

Jawed · 07-Jan-2008, 17:54

Quote:

Originally Posted by Geo

http://blogs.sun.com/HPC/entry/conso...idia_tesla_for

Is that part of the dump upstream or something new?

Now promising double-precision 2008H1.

Jawed

Tim Murray · 08-Jan-2008, 03:25

Quote:

Originally Posted by Jawed

Now promising double-precision 2008H1.

Jawed

More support for G90! More support for G90!

(okay I don't actually know anything real about this so give me a break)

Geo · 08-Jan-2008, 03:40

Well, y'know, VR-Zone claimed that G100 was coming in 1Q and we laughed at them. But they tend to be right more often than not with nvidia stuff, and ghu knows that "stupid naming tricks" have been rampant of late, so. . .

13-Nov-2007, 04:44	#1
Tim Murray whoops Join Date: May 2003 Location: Santa Clara, CA Posts: 3,266	NVIDIA CUDA dump from SC2007 Huge amount of stuff I'm still parsing, so you'll see some more news/real analysis soon--just wanted to get this out there first. - CUDA 1.1 beta available for registered developers--debugging on the GPU, visual profile, async support, future Windows drivers will have CUDA support in the main branch, stream management? Lots and lots and lots of stuff. - The eight hour CUDA presentation from SC07. Still going through it, but they're still saying DP next month. So far, I've read a lot of things that I've already seen, but give me some time. And of course, thanks to the lovely person who pointed this out to me

07-Jan-2008, 17:36	#8
Geo Mostly Harmless Join Date: Apr 2002 Location: Uffda-land Posts: 9,156	http://blogs.sun.com/HPC/entry/conso...idia_tesla_for Is that part of the dump upstream or something new? __________________ "We'll thrash them --absolutely thrash them."--Richard Huddy on Larrabee "Our multi-decade old 3D graphics rendering architecture that's based on a rasterization approach is no longer scalable and suitable for the demands of the future." --Pat Gelsinger, Intel ". . .its taking us longer than we would have liked to get a [Crossfire game] profiling system out there" --Terry Makedon, ATI, July 2006 "Christ, this is Beyond3D; just get rid of any fker talking about patterned chihuahuas! Can the dog write GLSL? No. Then it can fk off." --Da Boss

08-Jan-2008, 03:40	#11
Geo Mostly Harmless Join Date: Apr 2002 Location: Uffda-land Posts: 9,156	Well, y'know, VR-Zone claimed that G100 was coming in 1Q and we laughed at them. But they tend to be right more often than not with nvidia stuff, and ghu knows that "stupid naming tricks" have been rampant of late, so. . . __________________ "We'll thrash them --absolutely thrash them."--Richard Huddy on Larrabee "Our multi-decade old 3D graphics rendering architecture that's based on a rasterization approach is no longer scalable and suitable for the demands of the future." --Pat Gelsinger, Intel ". . .its taking us longer than we would have liked to get a [Crossfire game] profiling system out there" --Terry Makedon, ATI, July 2006 "Christ, this is Beyond3D; just get rid of any fker talking about patterned chihuahuas! Can the dog write GLSL? No. Then it can fk off." --Da Boss

Thread Tools
Show Printable Version Email this Page
Display Modes
Linear Mode Switch to Hybrid Mode Switch to Threaded Mode

13-Nov-2007, 04:59	#2
Rufus Member Join Date: Oct 2006 Posts: 214	And for those of us into a bit of theory wankery, there's a beta release of CUDPP (Cuda Data Parallel Primatives) up here: http://www.gpgpu.org/developer/cudpp/ It's a library consisting of scan, sort, and pack. Doesn't sound like much, but there is a huge class of parallel algorithms that depend on nothing but those 3, and if those 3 are done well the algorithms will Just Work.

13-Nov-2007, 05:50	#3
Tim Murray whoops Join Date: May 2003 Location: Santa Clara, CA Posts: 3,266	I just read the optimization presentation. My head hurts a lot.

13-Nov-2007, 08:20	#4
Lux_ Member Join Date: Sep 2005 Posts: 168	Last link is broken (Wave Propagation Through Earth)?

14-Nov-2007, 03:21	#6
cho Member Join Date: Feb 2002 Posts: 366	there is a paper about DP by University of Virginia and NVIDIA Research : http://www.cs.virginia.edu/~ktd3q/docs/ssr_gpu_ije.pdf "We show that SSR can be used effectively in programmable graphics architectures to allow double-precision computation without affecting the performance of single precision calculations and to increase fragment shader performance with a minimal impact on chip area."