SuperPi on GPU, were going CUDA

[-X-hellfire]

yeah i know, but it would great if these benchmarks did something useful, cure cancer or something while burning cpu/gpu cycles...


edit: but it's probably not possible to create a benchmark that uses "livedata" and then transmits results of it when its done and get comparable benchmark data, like superpi...

[bowman]

Well they do, just run WCG/BOINC or F@H and compare PPD numbers.

[-X-hellfire]

cool

why is then super pi etc mostly used to show/compare results?

[STEvil]

Because they're more reliable numbers than BOINC(WCG)/F@H can give.

Points are assigned per F@H Work Unit and they are not equal in value between Cell, ATi, nV, and CPU

BOINC may provide better numbers but nobody knows about optimizations etc so all we can do is wait. Also only nV cards are supported at the moment I believe.

[Marvin_The_Martian]

And will ever be. CUDA isn't licensed to anyone but Nvidia atm.

Using FAH wouldn't work because you will need the exact same wu / run / clone to run on all systems ( impossible ) and within other wu's there will always be a variation.

CUDA-Z is looking like the first usable cuda program for us

FUGGER any updates on a CUDApi?

[Shintai]

CUDA looks so useless after the Larrabee PDF...

[Marvin_The_Martian]

How do you mean? CUDA is present time, Larrabee isn't

Also CUDA is an API, Larrabee is an architecture

[Shintai]

How do you mean? CUDA is present time, Larrabee isn't

Also CUDA is an API, Larrabee is an architecture

And Larrabee is ANSI C/C++ and x86. CUDA is very limited in functionality. It will be a very short time before people say CUDA? What CUDA?

Larrabee will able to do folding and such several times faster due to no scattered writes and such.

Larrabee will also be able to handle ALL WUs. Not just a small minority like GPUs only can.

[Marvin_The_Martian]

True, but point one keeps up for now. When Larrabee arrives and it holds all that it's said to be I'll happily use it for folding

And CUDA doesn't seem that limited, it just has a niche market atm. When developers pick up on it, it might still be around in a decade in some form. Once a language is more used, the support base grows and the developers also commit more resources to it allowing them to maybe add new functionality.

I do find it a BIG TURN OFF that it's impossible to run CUDA.net on vs2008 vista x64.. allot of dissapointed people there, allot of talk how backwards Nvidia is with not supporting the most widely used ide in windows development for their 'cutting edge' technologie.

[MrHydes]

yeah i have here some screens of Cuda-z 3.50

By MrHydez

download Cuda-z

[bowman]

It is limited, the programmer is left with the burden of optimizing for each specific architecture, the compiler doesn't do much of that work. Getting a CUDA app to fully utilize the GPU is difficult. And then when a new GPU arrives you'll need to rewrite your app to utilize it fully because it's a new architecture again..

What Larrabee provides is not just a more widely used language but also better compilers, better scaling (less dedicated hardware) and forwards-compatibility. You just know that the Larrabee you buy today will still be relevant in ten years. A DirectX-based GPU with lots of fixed-function hardware and a niche language that will probably change so much for each new hardware iteration it's almost pointless? Lol..

Jen-Hsun Huang is right when he says that Larrabee is a dark Powerpoint slide cloud over them and their product. What he should do, instead of throwing a tantrum, is to do something about it and make his GPUs truly programmable. Give them a true ISA (doesn't even need to be x86, just widely accepted). Give us an actual roadmap. Right now the GPU manufacturers follow Microsoft in lock-step for each DX version. Larrabee completely invalidates that business model and forces them to adapt (thank goodness). Tim Sweeney said that DX10 would be the last relevant API, that software rendering would take over in this timespan going forwards, and I guess he was right.

[Marvin_The_Martian]

I thougt CUDA just present the programmer with a potential number of threads/blocks/memory which you need to read at run time to determine the possibilities of parrelisation your code? You can't have the compiler do that as your program would indeed be static. I'm sorry maybe I'm wrong but on that aspect I think CUDA isn't 'not scaling with hardware' it's not supposed to it's an API and it's only as good as the programmer is able to forsee what hardware it will be run on.

I'm just a hobbist though so probably I'm looking at this the wrong way?

[K3DK]

yeah i have here some screens of Cuda-z 3.50

hmm... i have 9600GT and very similar score:

CUDA-Z Report
=============
Version: 0.3.50
http://cuda-z.sourceforge.net/
OS Version: Windows x86 6.0.6001 Service Pack 1

Core Information
----------------
	Name: GeForce 9600 GT
	Compute Capability: 1.1
	Clock Rate: 1750 MHz
	Wrap Size: 32
	Regs Per Block: 8192
	Threads Per Block: 512
	Threads Dimentions: 512 x 512 x 64
	Grid Dimentions: 65535 x 65535 x 1

Memory Information
------------------
	Total Global: 512 MB
	Shared Per Block: 16 KB
	Pitch: 256 KB
	Total Constant: 64 KB
	Texture Alignment: 256
	GPU Overlap: No

Performance Information
-----------------------
Memory Copy
	Host Pinned to Device: 3079.15 MB/s
	Host Pageable to Device: 1275.36 MB/s
	Device to Host Pinned: 3090.34 MB/s
	Device to Host Pageable: 1327.33 MB/s
	Device to Device: 22320.7 MB/s
GPU Core Performance
	Float Point: 221938 Mflop/s
	32-bit Integer: 44680.8 Miop/s
	24-bit Integer: 220943 Miop/s

Generated: Mon Aug 18 01:23:39 2008

[Manabu]

Don't know about the open source part, but x264 encoder by cuda seems to be on.

Correction: there is no x264 encoder in cuda, only some crappy H.264 encoders in cuda, like badaboom.

bit more difficult than once thought....?

It seems so.

Dark Shikari (x264 developer) said: "CUDA being "just like standard C" is true, but only syntactically--the threading model is what makes it very difficult to use." , " After my experience trying to port something as trivial as an exhaustive motion search to CUDA, I'm not touching it with a ten foot pole."

Also, the undisclosed aparently high latencies dificulties even more get an high performance in GPU.

[Lestat]

It is limited, the programmer is left with the burden of optimizing for each specific architecture, the compiler doesn't do much of that work

oh i am sorry that the programmers have to do their fvcking jobs and not rely on a preprogrammed set of tools to do it for them.

get real, that comment is LAZY, and reflects the mentallity of LAZY programmers. P E R I O D.

CUDA is C,,, what the hell,, i suppose you'd rather have it in some retarded unix language, or bsd or wtf ever... cuda requires skill. something far far FAR too many programers DONT have. so they graduated some programming classes,,, yeah, that means a whole lot.

all this complaining about cuda, but i bet when you fold with your video card you aint friggin complaining as your pointless points climb to such a degree it would have taken years with a cpu to get that high.

sorry for the "in your face" comments, but you all know me,, i refuse to hold back, and that comment was pure bullsh/t.
no one said cuda was going to be simple, and NO ONE and i mean NO ONE outside of the intel dungeons know ANYTHING about this larabee stuff, its all a dream,, numbers on paper that dont mean jack squat until a material platform comes about.

and CUDA was designed to take advantage of current and future hardwares without the NEED to upgrade,, i dont think i have seen anything even remotely close to that from intel, its all an entire platform change.

i am not saying larabee isnt going to smoke CUDA or if its going to be a joke,, all i am saying is specualting on a platform thats nothing but paper right now is rediculous.

nvidia was the first to do what is being done, and just as their video card market is, they wont let someone else be top dog, not for long, so if someone does come out with a new programming tool or device (larabee) then i am quite confident nvidia will fight back with something better.

[Marvin_The_Martian]

oh i am sorry that the programmers have to do their fvcking jobs and not rely on a preprogrammed set of tools to do it for them.

get real, that comment is LAZY, and reflects the mentallity of LAZY programmers. P E R I O D.

CUDA is C,,, what the hell,, i suppose you'd rather have it in some retarded unix language, or bsd or wtf ever... cuda requires skill. something far far FAR too many programers DONT have. so they graduated some programming classes,,, yeah, that means a whole lot.

all this complaining about cuda, but i bet when you fold with your video card you aint friggin complaining as your pointless points climb to such a degree it would have taken years with a cpu to get that high.

Lestat,

I read more comments by phd's and ing's who seem to agree with joe sixpack me CUDA support is very lacking atm... but you're entiteld to your own opinion sure. Just don't go around accusing people of being lazy or unskilled when a tool which isn't even supported on vista/vs08. Talking about cutting edge eh

[spdycpu]

yeah i have here some screens of Cuda-z 3.50

Odd score for a GTX 280, any idea what is wrong? Here is my 8800GT at ~750 GPU, 2000 shader, 2300 memory:

CUDA-Z Report
=============
Version: 0.3.50
http://cuda-z.sourceforge.net/
OS Version: Windows x86 5.1.2600 Service Pack 2

Core Information
----------------
	Name: GeForce 8800 GT
	Compute Capability: 1.1
	Clock Rate: 1998 MHz
	Wrap Size: 32
	Regs Per Block: 8192
	Threads Per Block: 512
	Threads Dimentions: 512 x 512 x 64
	Grid Dimentions: 65535 x 65535 x 1

Memory Information
------------------
	Total Global: 511.688 MB
	Shared Per Block: 16 KB
	Pitch: 256 KB
	Total Constant: 64 KB
	Texture Alignment: 256
	GPU Overlap: Yes

Performance Information
-----------------------
Memory Copy
	Host Pinned to Device: 3179.55 MB/s
	Host Pageable to Device: 1690.55 MB/s
	Device to Host Pinned: 3179.76 MB/s
	Device to Host Pageable: 1699.44 MB/s
	Device to Device: 27607.3 MB/s
GPU Core Performance
	Float Point: 406304 Mflop/s
	32-bit Integer: 81650.6 Miop/s
	24-bit Integer: 406339 Miop/s

Generated: Wed Aug 20 05:48:25 2008

[gregorisvas]

Any update?

[Shadowmage]

Lestat,

I read more comments by phd's and ing's who seem to agree with joe sixpack me CUDA support is very lacking atm... but you're entiteld to your own opinion sure. Just don't go around accusing people of being lazy or unskilled when a tool which isn't even supported on vista/vs08. Talking about cutting edge eh

That's right. Ph.D interns at NV as well as Ph.D students working with CUDA at my university have explicitly told me that CUDA's functionality is currently very limited. It's only successful at obtaining "low-hanging fruit" in parallelism, and only supports a limited set of parallel programming models.

[Marvin_The_Martian]

That's right. Ph.D interns at NV as well as Ph.D students working with CUDA at my university have explicitly told me that CUDA's functionality is currently very limited. It's only successful at obtaining "low-hanging fruit" in parallelism, and only supports a limited set of parallel programming models.

You're actually promoting Lestats and Shintai's point more atm.

Maybe that's also because I defend CUDA since it is the base for the GPU folding which is currently revolutionizing how to simulate proteine folding. I know it's limited, maybe I defended it to much.

I said it's lacking in support, as it doesn't support the latest development tools not latest windows version. I can keep saying that.

I defended against it's limits in parallelism, maybe because atm, it's limited but it's all we get. I'n just happy it's here, right now, and it enabels us to do what we do with it. When all people do is trash it, I might respond out of emotion as well as reason but I still support everything I said.

[makaka]

yeah i have here some screens of Cuda-z 3.50

How come that the 8800 GT look faster on CUDA-z

[Warboy]

How come that the 8800 GT look faster on CUDA-z

it's called downclocking.

[makaka]

it's called downclocking.

could u post the actual screen shot of the GTX 280 @ default , the result look wired even if the GTX was down clocked

[Warboy]

Here

[MrHydes]

How come that the 8800 GT look faster on CUDA-z

yes, the GTX 280 in idle downclocks to 300MHz core, shader 602MHz and memory 1188MHz so when i posted the cuda z, wasn't in default so to speak.

cheers