19.11.2009 03:57:09 NVIDIA GPU 0: GeForce GTS 250 (driver version 19555, CUDA version 3000, compute capability 1.1, 512MB, 470 GFLOPS peak)

These numbers are strange...--a Are there new versions released for nVidia driver and CUDA?

I guess he use a BETA/prerelease nVIDIA_driver and CUDA_version.

If this would be the final release of 195.x and 3.0 , I guess we would hear it early enough.

But from that I heared, the 'old' GPU chips wouldn't get a (big) speed up with CUDA_V3.0 (like it was between 2.2 and 2.3).
CUDA_V3.0 would be well for the new 'Fermi' (GTX3xx, maybe also for the small brothers of) series GPUs.

Thanks for your explanation...:)

Actually it is a change in the Boinc 6.10.xx series that has changed how the benchmark is preformed. The 6.6.xx Boinc version would have shown it as something like this.

19.11.2009 03:57:09 CUDA device: GeForce GTS 250 (driver version 19555, compute capability 1.1, 512MB, est. 84GFLOPS)

Regards

---

Please consider a Donation to the Seti Project.

19.11.2009 03:57:09 NVIDIA GPU 0: GeForce GTS 250 (driver version 19555, CUDA version 3000, compute capability 1.1, 512MB, 470 GFLOPS peak)

These numbers are strange...--a Are there new versions released for nVidia driver and CUDA?

I guess he use a BETA/prerelease nVIDIA_driver and CUDA_version.

If this would be the final release of 195.x and 3.0 , I guess we would hear it early enough.

But from that I heared, the 'old' GPU chips wouldn't get a (big) speed up with CUDA_V3.0 (like it was between 2.2 and 2.3).
CUDA_V3.0 would be well for the new 'Fermi' (GTX3xx, maybe also for the small brothers of) series GPUs.

As always: why those guys who has CUDA 3.0 doesnot distribute dlls / suppose there is only 2 DLL like "old CUDA 2.3.
But no; they will always tell they cannot give anyone because it is prebeta ( or something like that)
It look like that two DLL.s are more secret than whole Pentagon :)

---

I am cruncher :)
I LOVE SETI BOINC :)

The 3.0 DLLs have been tested at Lunatics, and showed at best a 2% performance increase over 2.3 - nothing like the 30%+ achieved by 2.3 over 2.2, or 2.2 over 2.1. So nothing worth losing your licence over.

And losing your licence is important: if they deliberately flouted the developer NDAs (Non Disclosure Agreements), they could lose their access to pre-release development tools, or closed technical message boards where problems can be reported and overcome.

Apply for NVIDIA’s GPU Computing registered developer program like I replied to you in your asking for copies. Here's links:

http://www.nvidia.com/object/cuda_get.html
http://nvdeveloper.nvidia.com/content/GPUComputingDeveloperApplication/frmDeveloperRegistration.asp

---

...

The 3.0 DLLs have been tested at Lunatics, and showed at best a 2% performance increase over 2.3 - nothing like the 30%+ achieved by 2.3 over 2.2, or 2.2 over 2.1. So nothing worth losing your licence over.

And losing your licence is important: if they deliberately flouted the developer NDAs (Non Disclosure Agreements), they could lose their access to pre-release development tools, or closed technical message boards where problems can be reported and overcome.

Many of them are on that forum only for CUDA toolkit, and even loosing license they will not lose anything important: because that persons are not developer or something like that: they are people like ordinary one: just have lucky and get access to nvidia forum :)

---

I am cruncher :)
I LOVE SETI BOINC :)

Many of them are on that forum only for CUDA toolkit, and even loosing license they will not lose anything important: because that persons are not developer or something like that: they are people like ordinary one: just have lucky and get access to nvidia forum :)

If you can write code like Jason or Raistmer then get stuck in - we still need to solve the VLAR problem.

Otherwise, there are clearly two grades of 'ordinary' people.

The 3.0 DLLs have been tested at Lunatics, and showed at best a 2% performance increase over 2.3 - nothing like the 30%+ achieved by 2.3 over 2.2, or 2.2 over 2.1. So nothing worth losing your licence over.

And losing your licence is important:....

so when they tested 3.0 DLLS on Lunatics ,that is ok ,and if they put on Internet that is not ok :)
How interesting :)

---

I am cruncher :)
I LOVE SETI BOINC :)

so when they tested 3.0 DLLS on Lunatics ,that is ok ,and if they put on Internet that is not ok :)
How interesting :)

Testing is an important part of development (as you must know, from your 'Volunteer tester' tag).

And believe me, if those tests had revealed another 30% increase, they would have been falling over themselves to develop a safe, reliable, tested solution to use the 3.0 DLLs for SETI (ensuring that the tasks validate, and hence generate credit, when compared with older versions) - while still keeping their licences to develop with 3.1, 3.2,4.0, ...

"CUDA device: GeForce GTX 275 (driver version 18585, compute capability 1.3, 896MB, est. 123GFLOPS)

CUDA device: GeForce GTX 285 (driver version 18171, CUDA version 1.3, 1024MB, est. 127GFLOPS)"

Was wondering if this is a good comparison? Seems for the money the GTX 275 is almost as fast as the GTX 285. Is there a big difference between Gddr3 and ddr3 memory? They both seem to have 240 processors. Anyone have both?

---

Official Abuser of Boinc Buttons...
And no good credit hound!

[http://www.nvidia.com/object/cuda_learn_products.html]
[http://www.nvidia.com/object/product_geforce_gtx_275_us.html]
[http://www.nvidia.com/object/product_geforce_gtx_285_us.html]

GTX275 - 633/1404/1134 - (123 GFLOPS ?)
GTX285 - 648/1476/1242 - (127 GFLOPS ?)

[GPU/shader/RAM]

The upper mentioned GPUs are at stock speed?

My OCed GTX260-216 have:
EVGA GTX260 Core216 SSC - 675/1458/1152 -> 112 GFLOPS
GIGABYTE GTX260(-216) SOC - 680/1500/1250 -> 117 GFLOPS
The GIGABYTE have higher speeds as a GTX285!
But O.K., 30 (240) or 27 (216) CUDA (shader) cores.

Because of CUDA..
more shader cores
more shader speed
more RAM speed
more GPU core speed (maybe)
-> more performance

---

31/10/2009 09:07:21 NVIDIA GPU 0: GeForce 8400 GS (driver version 19062, CUDA version 2030, compute capability 1.1, 256MB, 43 GFLOPS peak)

to

27/11/2009 23:45:24 NVIDIA GPU 0: GeForce 8400 GS (driver version 19562, CUDA version 3000, compute capability 1.1, 256MB, 43 GFLOPS peak)


@+

BOINC V6.6.x

Manufacturer OCed

EVGA GTX260 Core216 SSC - 675/1458/1152 -> 112 GFLOPS
GIGABYTE GTX260(-216) SOC - 680/1500/1250 -> 117 GFLOPS

[GPU/shader/RAM]

A stock GTX260-216 -> 576/1242/999
[...]

After update to BOINC V6.10.18:

The EVGA brothers :
NVIDIA GPU 0: GeForce GTX 260 (driver version 19038, CUDA version 2030, compute capability 1.3, 896MB, 630 GFLOPS peak)
NVIDIA GPU 1: GeForce GTX 260 (driver version 19038, CUDA version 2030, compute capability 1.3, 896MB, 630 GFLOPS peak)
NVIDIA GPU 2: GeForce GTX 260 (driver version 19038, CUDA version 2030, compute capability 1.3, 896MB, 630 GFLOPS peak)
NVIDIA GPU 3: GeForce GTX 260 (driver version 19038, CUDA version 2030, compute capability 1.3, 896MB, 630 GFLOPS peak)

The GIGABYTE:
NVIDIA GPU 0: GeForce GTX 260 (driver version 19038, CUDA version 2030, compute capability 1.3, 896MB, 653 GFLOPS peak)

---

NVIDIA GPU 0: GeForce 8800 GTS 512 (driver version 19562, CUDA version 3000, compute capability 1.1, 512MB, 442 GFLOPS peak)

Not that I use my Graphics card for crunching, but I'll post here anyway.

28/11/2009 3:17:45 p.m.		NVIDIA GPU 0: GeForce 8600M GT (driver version 18681, CUDA version 2020, compute capability 1.1, 256MB, 61 GFLOPS peak)

Hopefully the GTX260 C216 Superclock in my gaming build will do a bit better!

---

- Luke.

I'm curious,
How many GFLOPS can an Intel Core 2 Quad Q6600 do at stock speed?

---

I'm curious,
How many GFLOPS can an Intel Core 2 Quad Q6600 do at stock speed?

About 2.4 per core, or 9.6 in total. Floating Point speed is very close to stock clock for the Core 2s.

I'm curious,
How many GFLOPS can an Intel Core 2 Quad Q6600 do at stock speed?

About 2.4 per core, or 9.6 in total. Floating Point speed is very close to stock clock for the Core 2s.

For the Core 2 architecture, each core has multiple execution units, one of which can do 4 SIMD floating point operations per clock, another can do one floating point operation per clock. So 2.4 * 5 = 12 peak GFLOPS per core in terms comparable to the way the GPUs are rated.

                                                                 Joe

NVIDIA GPU 0: GeForce 9600 GT (driver version 19107, CUDA version 2030, compute capability 1.1, 512MB, 208 GFLOPS peak

---

For the Core 2 architecture, each core has multiple execution units, one of which can do 4 SIMD floating point operations per clock, another can do one floating point operation per clock. So 2.4 * 5 = 12 peak GFLOPS per core in terms comparable to the way the GPUs are rated.

                                                                 Joe

Looking for a direct comparison between the old and new BOINC estimates, I found Martin's GTS 250:

CUDA device: GeForce GTS 250 (driver version 0, CUDA version 1.1, 512MB, est. 84GFLOPS)
NVIDIA GPU 0: GeForce GTS 250 (driver version unknown, CUDA version 2030, compute capability 1.1, 512MB, 470 GFLOPS peak)

I reckon my 'est. 2.4GFLOPS' (benchmark per core) is pretty close to your '12 GFLOPS peak': it's just they way we count 'em.