I expect a cut down version of the GTX680 to arrive, probably a GTX670 and ~20% slower (192 less shaders). A GTX660Ti could possibly turn up too (384 fewer shaders than a GTX680).

While I suppose a GTX690 (dual GTX680) could arrive, I'm not so sure it's going to happen just yet.

Two smaller OEM cards have already arrived, both GT 640's (28nm, GK107). One is a GDDR5 version and the other uses DDR3.

I see the manufacturers have started launching non-reference GTX680 designs.
Galaxy is launching a non-reference, full GTX680 Kepler,


Typically, these come with redesigned boards, higher stock clocks, dual or triple fans, and some contain additional functionality (support for more monitors, more memory...). While 1202/1267MHz, for example, is a nice boost (19%) it's worth noting that some such cards need two 8-pin power connectors to power them.
____________
FAQ's

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

Find it weird that they release this card, when they can't even keep up with supplies for 680

So the GTX690 is being released on the 3rd May.

Price tag aside, I'm quite impressed with this.
Firstly, I wasn't expecting it just as soon, given the GPU shortages. This is the first time (in recent years anyway) that NVidia has released a dual GPU in a series ahead of AMD/ATI.

Two hardware specs are of obvious interest:
The Power requirement is only 300W. Considering a GTX590 requires 375W (TDP) that's a sizable drop. This leaves the door open for OC'ing and might mean there is less chance of cooking the motherboard.
The other interesting thing is the GPU frequency, relative to the GTX680. Both cores are at ~91% of a GTX680. A GTX590's cores were only 78% of a reference GTX580 - which made it more like two GTX570's in terms of performance.
This certainly looks like a good replacement for the GTX590 (if you don't need FP64).

There seems to be a noticeable improvement in the basic GPU design, but I have not yet seen a review of actual performance (noise, heat, temps).
____________
FAQ's

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

0b1010110010 =))
____________
FAQ's

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

The GTX680(or future GTX690)work well and without problems GPUGRID ...? Thanks

It's the same chip with similar clocks, but less shaders active -> less performance.

MrS
____________
Scanning for our furry friends since Jan 2002

Quite the opposite actually!

A GTX570 was ~89% as fast as a GTX580.

The difference in terms of performance between the GTX670 and GTX680 will be ~20% :
1344/1536 shaders * 915/1006 MHz = 0.796 (~80%).

Assumes performance can be calculated in the same way as previous generations...

The GTX580 only had 512 shaders, so disabling just 32 was not a lot.
The GTX680 has three times the shaders (1536), but disabling 192 is relatively twice as much; 32*3 would only be 96.

So I'm hoping for a price closer to £300. When you consider that the HD7970 presently costs ~£370 and the HD7950 can be purchased for as little as £299, the GTX670 is likely to come in at around £330. While this is similar to original GTX470 and GTX570 prices, it would be better value at £300 to £320.
To begin with the lack of supplies will justify a £330 price tag, but unless it beats an HD7970 in most games (unlikely) anything more than £330 would be a rip-off - when it does turn up the HD7950 and HD7970 prices are likely to fall again.

At present most UK suppliers don't have any GTX680's, and are listing them as £400 to £420 (stock delayed, 3-4 weeks, down to 4days).
I did see a GTX690 at £880 (5-7days), but at that price what you save in terms of performance per Watt would be lost in the original price tag.

Anyway, £400 * 0.8 = £320.
When the GTX670 does arrive, I would not expect its price to fall for a couple of months, or at least until GTX660's turn up (two of which may or more likely may not be a viable alternative).
No idea when AMD/ATI will have their next refresh, but it's likely to be a while (few months), and might depend more on the new memory than anything.
If anything it will probably be competition between the rival manufacturers that drives prices down (and up), with various bespoke implementations.
____________
FAQ's

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

so would it be "safe" to say a 670>580 here? Based on the #'s?

A 1.5GB HD7950 would be a good card to throw into the mix. Price likely to be ~£250 to £280, so very competitive. It would also go some way to filling the gap between the HD7950 and HD7870, and compete against any forthcoming GTX 660Ti/660 which are anticipated to arrive with 1.5GB.

All good news for crunching.
____________
FAQ's

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

The only BOINC project I know which currently uses DP is Milkyway.. where the AMDs are superior anyway.

MrS
____________
Scanning for our furry friends since Jan 2002

I agree, neither a Fermi or a Kepler is any good at FP64. It's just that Kepler is relatively less good. Neither should be considered as practical for MW. Buying a Fermi rather than a Kepler is just sitting on the fence. If you want to contribute to MW get one of the recommended AMD cards. Even a cheap second hand HD 5850 would outperform a GTX580 by a long way. If you want to contribute here get one of the recommended cards for here. As yet these don't include Kepler, because I have not updated the list, and won't until there is a live app. That said, we all know the high end GTX690, GTX680 and GTX670's will eventually be the best performers.
____________
FAQ's

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

The Folding@Home CUDA benchmark is interesting, as it has a reference GTX670 within 6% of a reference GTX680.

Anandtech, http://images.anandtech.com/graphs/graph5818/46450.png

Despite using CUDA this app is very different to the one here and it would be wrong to think such performance is likely here.

Going by the numerous manufacturer's implementations I'm now expecting substantial variations in performance from GTX670 cards; from what I can see the boost speeds will tend upwards from the reference rates (980MHz), with some FOC's reaching ~1200MHz (+22%), which would best a reference GTX680. Overclockers have went past this and 1250 or 1266 is common. Some even reference >1300MHz on air. Of course a GTX680 can overclock too. For hear the OC has to be sustainable, stable for weeks/months rather than minutes.
____________
FAQ's

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

This is a little off-topic, but I'll add my 2 cents anyway: bying anything AMd older than the 5000 series wouldn't make sense since these 55+ nm chips are less power efficient. Currently the best bet would probably be a Cayman chip. An unlocked HD6950 is still very good value.

At MW my Cayman, which I bought over 1 year ago for a bit over 200€, was doing a WU each ~50s. I guess that's tough to beat for a GTX580 ;)

Anyway, it's running POEM now: more credits, less power consumption, hopefully more useful science :)

MrS
____________
Scanning for our furry friends since Jan 2002

I don't have new hard numbers, but Kepler is looking quite good for GPU-Grid. Comparing GTX580 and GTX670 you'll save almost 100 W in typical gaming conditions. At 0.23 €/kWh that's 200€ less of electricity cost running 24/7 for a single year. I can't say how much the cards will draw running GPU-Grid, but I'd expect about this difference, maybe a bit less (since overall power draw may be smaller).

That would mean choosing a GTX670 over a GTX580 would pay for itself in about 6 months.. pretty impressive!

MrS
____________
Scanning for our furry friends since Jan 2002

A reference GTX580 has a TDP of 244W, and a GTX670 has a TDP of 170W.
GPUGrid apps typically use ~75% of the TDP. So a GTX580 should use around 183W and a GTX670 should use around 128W. The GTX670 would therefor save you ~55W.
____________
FAQ's

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

You're assuming Keplers use the same percentage of their TDP-rating as Fermis. That may not be true for several reasons:

- different architecture
- on high end GPUs nVidia tends to specify unrealistically low TDPs (compared to lower end parts of the same generation)
- for Keplers the actual power consumption is much more in line due to their Turbo mode, this essentially makes them use approximately their maximum boost target power under typical load

For the 100 W difference in games I'm referring to such measurements.

MrS
____________
Scanning for our furry friends since Jan 2002

Yes, but this is untested. My guess ~4.5%
AMD don't support PCIE3.0 Motherboards/CPU's!!
Also, worth noting again that there is no XP support for GTX600's.
For a single card I'm thinking IB is the best option, but outlay for such a rig is very high, and I'm convinced that moving from SB to IB is not a realistic upgrade; it's more of a detour!
As AMD offer much cheaper systems they make good sense. >90% as efficient at half the outlay price makes sense. Running costs is open to debate, especially if you replace often.

So, for a single card (especially GTX670 or less) don't let PCIE2 deter you.
____________
FAQ's

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

View PCIe 3 as a bonus if you get it, but noting else. At least for GPU-Grid.

For a single card I'm thinking IB is the best option, but outlay for such a rig is very high, and I'm convinced that moving from SB to IB is not a realistic upgrade; it's more of a detour!

It means that Folding@Home is a different CUDA project. Which means that the relative performance of cards with different architectures may not be the same on GPU-Grid.

Anyway, the app supporting Kepler should be out in a few days and we'll have hard numbers. So far GTX670 looks pretty good here.

MrS
____________
Scanning for our furry friends since Jan 2002

Nice specs from VideoCards, albeit speculative:

GeForce 600 Series GPU CUDA Cores Memory Memory Interface Launch Price Release
GeForce GT 640 D3 GK107 384 1GB / 2GB GDDR3 128-bit $99 April
GeForce GT 640 D5* GK107 384 1GB / 2GB GDDR5 128-bit $120 August
GeForce GT 650 Ti* GK106 960 1GB / 2GB GDDR5 192-bit $120 – $160 August
GeForce GTX 660* GK106 1152 1.5 GB GDDR5 192-bit $200 – $250 August
GeForce GTX 660 Ti* GK104 1344 1.5GB / 3GB GDDR5 192-bit $300 – $320 August
GeForce GTX 670 GK104 1344 2GB / 4GB GDDR5 256-bit $399 May
GeForce GTX 680 GK104 1535 2GB / 4GB GDDR5 256-bit $499 March
GeForce GTX 690 2xGK104 3072 4GB GDDR5 512-bit $999 May

*Not yet released
http://videocardz.com/33814/nvidia-readies-geforce-gtsx-650-ti-for-august

The 660Ti might be a sweet card for here, if the specs are true, and we would have a good mid-range to choose from. No mention of a 768 card though, and there is a bit of a gap between 960 and 384. Their imagination doesn't seem to stretch down to a GT 650 (no Ti, 768). I can't see a GT650Ti going for the same as a GT640 either, and if a 960 card cost $120, you could just buy 2 and outperform a GTX680 (all else being equal) and save ~$260. With those specs 3 GT650Ti's would almost match a GTX690, but cost ~half as much.
____________
FAQ's

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