I am looking at getting a Palit GTX670 to replace a GTX570. More specifically their Jetstream version. Has anyone tried a GTX670 on here? What are your crunch times like?

Overview
Memory: 2048MB / 256bit GDDR5
Clock : Base Clock 1006MHz/Boost Clock 1084MHz / 3054MHz (DDR 6108MHz
HDMI / DVI x2 / Display Port

Link to it here

---

BOINC blog

GTX 670, GTX 680, GTX 690 is not supported yet (by the CUDA 3.1 application). They will be supported (hopefully) in a couple of days by the CUDA 4.2 application, which is in beta testing right now.

On the first betas, my 670 was about 10% slower. Well find out the real differences soon enough though. This was on W7.

On the first betas, my 670 was about 10% slower. Well find out the real differences soon enough though. This was on W7.

Did things improve with the cuda42 version?

---

BOINC blog

I only ran the first beta set on the 670. I may have been playing D3 on the second round on that card.........

Wouldn't see how it would make a difference though. It's going to be slower than the 680 for sure, and -10% is pretty good considering the 680 is $100+ more. Not to mention a pain to find still.

Mark might have understood you said "GTX670 10% slower than his GTX570" rather than "GTX670 10% slower than GTX680".

MrS

---

Scanning for our furry friends since Jan 2002

Ah......

Sorry about that. Was typing quickly on my phone. Apologies.

Yes, my 670 was -10% slower than my 680. This was clock for clock. Set both the same speeds.

MarkJ, I think that looks like a decent GPU. The GTX670 should prove to be a good replacement for a GTX570. Perhaps around 30% more work per day than the GTX570 and for less energy, so ball park ~50% to 60% more efficient per Watt.

A GTX670 on W7 should ~match a GTX580 on WinXP.
In terms of performance per Watt however the GTX670 >> GTX500 series cards.

---

FAQ's

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

My GTX670 in Win 7 with 301.42 driver.

GPUgrid -
ACEMD beta version 6.43(cuda42) - GPU load 83% (NVIDIA Inspector 1.9.6.5)

Primegrid -
Genefer (WR) 1.07 (cuda32_13) - GPU load 99% (NVIDIA Inspector 1.9.6.5)

At GPUGrid, W7 and Vista suffer an 11%+ loss in performance compared to WinXP or Linux. When I tested a 2008 server the loss was only ~3%.

---

FAQ's

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

True enough. Application programming is the difference. All projets need to define their resources for programming based on the assets and abilities to accommodate the new tech. Please don't let these limitations confuse donors into thinking that a poject doesn't care in the "short run". :)
If it continues over a couple months... that's different :)
They appreciate us until proven differently.
EDIT: We all have to remember that vista and W7 are an entirely different OS background. All of the pogramming has to be adjusted for V and W7. It's not Just the NV or AMD drivers. That's why NV has separate drivers for XP and previous versions and Vista and W7. The latter are not bases on NT tech ;)
Vista never took "massive hold" and neither has W7 yet. Programmers have to do what best results in results.
JIMO, YMMV (Just in my opinion, your mileage may vary) :)
sean

Vista and 7 sure enough are based on the NT code basis. Win 7 still identifies itself as ver 6.1., which is referring to the old NT nomenclature. The difference you're talking about is the display driver model, which changed with Vista.

MrS

---

Scanning for our furry friends since Jan 2002

Back to the original question...

I have ordered 2 of them. The supplier has a limit of 2 per customer anyway :-)

I see the "production" apps haven't switched to cuda40 or cuda42 yet, so will have to leave the GTX570's in place until that happens. I need to sell the old cards to help pay for the new ones. Hopefully it will happen fairly soon.

---

BOINC blog

Back to the original question...

I see the "production" apps haven't switched to cuda40 or cuda42 yet, so will have to leave the GTX570's in place until that happens.

The short queue already have the cuda4.2 application, and a couple of GTX 680s and GTX 670s are already crunching "production" tasks.

What do GPUGrid pros make of this statement from Anandtech regarding the 690?

"Unfortunately for NVIDIA GK104 shows its colors here as a compute-weak GPU, and even with two of them we’re nowhere close to one 7970, let alone the monster that is two. If you’re looking at doing serious GPGPU compute work, you should be looking at Fermi, Tahiti, or the future Big Kepler."

http://www.anandtech.com/show/5805/nvidia-geforce-gtx-690-review-ultra-expensive-ultra-rare-ultra-fast/15

Removing power consumption from consideration, does this really suggest that 590 is preferable to a 690 for GPUgrid? Asked differently, what can one extrapolate from that review/benchamark about a 690's performance here?

They're discussing double precision. This project is single precision.

Well one of them is installed and off and running. Pretty pics can be found here

Fortunately it picked up a cuda42 work unit to start with, an IBUCH TRYP, but only seems to be using about 48% load (peak 55%), so its hardly stressing the card.

---

BOINC blog

There other ibuch tasks around that get up to 96. The one you have now are actually the slowest out of all the different WUs in the short queue.

Asked differently, what can one extrapolate from that review/benchamark about a 690's performance here?

Easy: multiply the throughput of a GTX680 by 2 and you're basically there. Average clock speeds will be slightly lower, but this approximation should be good enough.

Otherwise.. as 5Pot said: DP performance is irrelevant here :)

MrS

---

Scanning for our furry friends since Jan 2002

I have a MSI factory overclocked 670 and it gives me computation errors in Linux right off the bat with the 302.17 drivers. I'm thinking it has something to do with the overclock since other people seem to be crunching just fine on Linux with the newest drivers and a 670. Either way, I've reverted back to 295.59 and it averages about four hours on a cuda42 long run, which I would assume is decent since it says 8-12 hours on fastest card in the description.