Either way, I've reverted back to 295.59 and it averages about four hours on a cuda42 long run

... which tells us that the 302 driver was to blame, not the factory overclock. Bad for nVidia, good for you :)

MrS

---

Scanning for our furry friends since Jan 2002

HI: I also have installed the 302.17 and with my GTX295 there is no way to run CUDA 4.2 tasks, from what I read is a matter of reinstalling the old 295.59 as soon as you finish the tasks that I have running CUDA 3.1... We'll see.

I have a GTX480 and it runs GPUGrid Fine and Dandy under the Linux 302.17 driver.

So much so, I am approaching the top 100 GPUGrid clients based on RAC (actually no 146 earlier today, up from 164 last night)

it runs CUDA31 and CUDA42 apps. Most CUDA31 Long runs take about 9 hours. CUDA42 Long runs take about 5 hours.

I had been running some PrimeGrid Cuda WUs until I found out that most Cuda work there is Double Precision. I didn't realize 680s were worse at DP than the 500s are. I guess SP projects are the way to go for the 600s series. I did get the 680 signature 2 with 2 fans and its very fast at SP work such as GPU Grid. I guess Milky Way would be slower also. Its kind of dissapointing Nvidia handicapped the 600s in DP. Maybe the GK110 will be better.

Yes, you can expect GK110 to be a DP monster.

MrS

---

Scanning for our furry friends since Jan 2002

Thats great, do you know if most of the projects will all eventually be DP, or is it only the Mathmatically leaning projects that will be this way.

My 570 is faster than my 680 by 15-20% on DP work.
GK110 is on my list definitly.
I just hope its in the less than 600-700 dollars though. That hurts.
I may be released as Tesla or Quadro only, we will see I suppose.

Each project has different criteria. GPUGrid does not need FP64/DP.

The GTX 680 is much better for here. While the 570 is still good, it's not great at DP. It's just that the GTX680 is awful. If you want a top DP card get a high end AMD/ATI card such as an HD7970. The GeForce cards are by in large better for SP.

A GeForce GK110 will be very pricey, and won't be available this year. If and when a 'GTX 685' does turn up, I would speculatively expect a performance increase of ~70 to 85%, over the GTX680 for here. On DP projects it would obviously be massive compared to the GTX680. Big Kepler will arrive in the form of a Tesla K20 some time this year, probably Q4, but you won't be buying one!

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

do you know if most of the projects will all eventually be DP

Never. And that's a good thing :)
The point being: it always takes more energy and hardware to do DP calculations. And it's not hard to design your DP hardware so that it can do 2 SP operations instead of 1 DP. So at best you can do DP at half the SP rate. That's why in performance-critical applications you should use SP when ever the precision is sufficient.

MrS

---

Scanning for our furry friends since Jan 2002

I think most of you ignored the fact that Anand's compute benchmark consisted of mostly if not all SP benchmarks.

It does seem like Kepler is a step backwards in terms of GPGPU.

Well, for here it's a good step forward. Obviously it's not for MW and some other projects. Pick a project and pick your cards for it. If you want to run POEM or MW get AMD cards. For here and Einstein get NVidia cards.
The bottom line is that these are gaming cards. It just happens to work well here because it supports CUDA and GPUGrid doesn't need DP. Architecturally there are still issues with it for crunching here but performance is still better than the previous generation.
If you want a full-fat compute card from NVidia you will have to wait for the Tesla K20 (Q4) and then you'll have to be prepared to pay for it! It's unlikely that a GeForce variant of this will appear until next year.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

GP-GPU does not equal DP crunching. In fact, even with the CC 3.0 Keplers being a step backwards in DP performance, this doesn't really matter, since AMD is far superior in raw DP performance to any Fermi or earlier...

MrS

---

Scanning for our furry friends since Jan 2002

4 hours is good.

Most of my CUDA42 workunits take about 5.5 hours on my GTX480

That is a bit shy of a 30% performance improvement for double the price.

Course power consumption drops by a similar amount. Its hard to factor the power consumption figure into the overall household utility expense in most areas of the USA (where time/demand pricing is not in force - time of day consumption may raise the utility cost by 100% or more compared to other times of the day

Its hard to factor the power consumption figure into the overall household utility expense in most areas of the USA

That's true, but you don't have to do it. 1$ electricity cost is 1$, no matter how much your other devices consume. All you really need is to measure the power consumption at the wall with and without GPU-Grid (or PC on/off) and multiply by running time and your local cost per electricity.

MrS

---

Scanning for our furry friends since Jan 2002

Hello, is it a good idea by following:

I want to overclock my GTX 670 (Asus)

Now i thought i could make it easy, an so:

Raise the voltage by 1mV per 1 Mhz, is it a good idea
or will i burn my chipset away !?

Greets :)

Bad idea. First: increase clock speed without voltage increase and see how far you get (should be somewhere around 1050 - 1100 MHz, from what I've read). If you're comfortable with the temperature, power consumption and noise at this setting you can push further. At that point "1% more voltage for 1% higher clock" is a fair approximation, although the real function is at least quadratic, maybe even exponential.

MrS

---

Scanning for our furry friends since Jan 2002

hm, i dont did it. Increasing the "power target" does it all. It clocks itself. no thinking about overvolting. Thanks.

hm, i dont did it. Increasing the "power target" does it all. It clocks itself. no thinking about overvolting. Thanks.

What target have you set? So Might do it as well.

On my GTX 670 it does not matter what I set the power target at it always pulls 1175. So knowing that I was not able to increase / decrease volts I might as well OC as far as is stable ... turns out that 1259 GPU and 3206 MEM is rock solid stable - 99 consecutive successful LONG WUs so far. Win7x64, BOINC 7.0.25.

Looked through some results for GTX 670s accross GPUGrid for Win7x64 with NATHAN WUs: OC I listed above takes about 9%-10% less time per WU and gets me to the ballpark of a stock GTX680!!!

---

Thanks - Steve

i did base overclocking (but to far) , now at 915+90mhz and set power target with Nvidia Inspector at 122% but it wont raise further that 1175mV. Not always stable right now.

I've begun to upgrade my Fermi cards to Kepler cards, and I've made some measurements with my first partly upgraded system.
Its configuration is (at the moment):

ASUS P7P55 WS Supercomputer motherboard
Intel Core i7-870 @ 3855MHz (24*160MHz)
2*2Gb DDR3 2000MHz RAM
320GB HDD
MSI GTX 480 @ 800MHz (1.075V) (with an Artctic Cooling Accelero Xtreme Plus)
Asus GTX-670 DC2 @ 1084MHz (1.137V) (factory overclocked)

A NATHAN_RPS_1120528 runs nearly 13% faster (15137 sec vs 17133 sec) on the GTX 670 @ 1084MHz than on the GTX 480 @ 800MHz.
I've also measured the power consumption at the wall outlet (230V AC).
When my PC was idle (no tasks running, but power management is disabled, so the CPU runs at full speed) it is drawing 178 Watts
When a task is running on the GTX 480 @ 800MHz (99% GPU usage): 378 Watts
When one more task is running on the GTX 670 @ 1084MHz (99% GPU usage): 552 Watts
When 4 rosetta@home are running on the CPU: 625 Watts
So, the extra power consumption of the different parts when they are in use is:
GTX 480 @ 800MHz (99% GPU usage): 200 Watts
GTX 670 @ 1083MHz (99% GPU usage): 174 Watts
CPU 4 cores: 73 Watts
As you can see, the GTX 670 @ 1083MHz consumes 87% of the GTX 480 @ 800MHz.
Hopefully I can do more measurements in this week.