http://www.anandtech.com/show/5911/nvidia-announces-retail-gt-640-ddr3,
by Ryan Smith.

I'm just saying they now exists, and I'm not saying 'buy one'!
____________
FAQ's

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

Should work in the same way as the bigger Kepler cards. Expect ~1/4 the performance of a full GK104 at similar clocks.

MrS
____________
Scanning for our furry friends since Jan 2002

12.24 Gflops per wattage = means 65 * 12.24 = ~ 800 gflops , not bad :)

I expect the GT 640 should just about match a GTX 550 Ti in terms of runtime performance, but maybe the 4.2 app will be favourable to it. Obviously the 65W TDP is much more attractive than the 116W of the GTX550Ti.
At the end of 2009 the GT240 entered the scene. It was a bit of a sweet entry level card; an early 40nm GPU only requiring power from the PCIE slot (69W TDP). If the GT 640 offers reasonable performance at <75W then it might be a good entry level card and allow more people to contribute. The recommended minimum PSU for such cards is only ~350W. I would also expect the GT640 to actually use about 35 to 40W.
____________
FAQ's

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

Performance roughly as expected. It would probably match the GTX 550 Ti, if it could run the CUDA 3.1 app, but on the CUDA 4.2 app it's actually matching a GTX460, due to the GTX600 series improvements in performance over the GTX400/500 series. Cool.
____________
FAQ's

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

Why must it draw 65 W? Because TDP is stated as 65 W by nVidia? Well.. no. Cards usually draw less than TDP at GPU-Grid and crunching in general. Consider this: you're probably not even at 100% GPU utilization, let alone using the TMUs etc.

So it seems like such a card is good for ~44k RAC here.

Overclocking actually improves power efficiency here. Overvolting doesn't.. but I hope you're not pushing the card to 2x the power consumption ;)

MrS
____________
Scanning for our furry friends since Jan 2002

Long term overvolting is even less recommended on lesser cards.
These are fairly 'green' cards compared to previous generations, but the bigger cards probably do more per watt, especially if you look at the power consumption of the system. Their niche is low end desktop systems with average PSU's.
As it's matching a GTX460 then it's doing the same work for half the electric consumed by the GPU.
The ~44k is for normal tasks. This should improve for the longer tasks, especially if the card can return them inside 24h.
____________
FAQ's

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

Well.. overvolting it still wouldn't cost much.

Stock load voltage seems to be 1.00 V. AMD is using 1.20 stock on the same process, so this would be the maximum I'd think about using. At 1.10 V power consumption would increase by a factor of 1.21 (1.10^2/1.00^2), pushing the card from an estimated 50 W to 60.5 W and maybe getting you another 100+ MHz core clock (-> ~10% performance increase).

Overall that's not too bad (considering you'Ve got an entire PC up'n running to feed the card), but 1.20 V may push the boundaries of PCIe power delivery: 1.20^2/1^2 = 1.44, i.e. a 44% increase in power consumption. Here we'd be talking about 72 W for maybe another 50+ MHz - diminishing returns.

MrS
____________
Scanning for our furry friends since Jan 2002

MIS Kombustor perhaps?
I think there are issues overvolting and possibly overclocking the GF600 cards. The software will turn up eventually, if there isn't already an overclocking kit. I don't have a GF 600 card so I haven't really looked into this.
____________
FAQ's

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

Apparently Voltage auto-adjusts on reference cards, so you can't manually do it, but it should increase with the GPU clock.
This change makes OC'ing the GF600's more like OC'ing the i7-2600K and similar CPU's which automatically adjust the Voltage as required.
____________
FAQ's

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

If it causes errors then don't increase the clocks.
It's possible that the reported Voltage is just a reference one, read off the cards reference Voltage or the driver, and not a real live reading.
____________
FAQ's

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

Which clock speeds do you reach?

Automatic voltage scaling is linked to turbo mode, which GT640 doesn't support.

MrS
____________
Scanning for our furry friends since Jan 2002

Side note: 100 watts per hour -> 100 W. Power is already divided by time.

MrS
____________
Scanning for our furry friends since Jan 2002

Each hour you save 0.1 kWh :)

MrS
____________
Scanning for our furry friends since Jan 2002

The GT640 should work in the second slot, though you might want to check your PCIE slot performance - if it's only PCIE x1 then if it even works it would be very poor. If both slots are PCIEx16 and stay X16 when the second slot is occupied then that's optimal. If the first slot drops to X8 and the second is only X4, then you would probably still do more work, but would be loosing a fair bit of performance through bandwidth limitations.

We really need to measure the performance loss when using PCIE3 vs PCIE2 and at various amounts of lane count. It was done roughly on PCIE2 with the old app for previous generations of cards, but things have changed since then. Some apps perform much better with additional PCIE bandwidth, especially PCIE3 as it's twice as fast as PCIE2.
____________
FAQ's

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

@klepel: the GTX560 xxxCore Edition is quite fast at GPU-Grid, way faster than the GT640. But it's also a totally different beast power-consumption-wise. Could your power supply, case cooling, ears and electricity bill stand such a card? IMO that would be the main question. GT640 is completely tame in comparison.

MrS
____________
Scanning for our furry friends since Jan 2002

A reference GTX560 Ti 448 is almost twice as fast as a GT640; ~1.85, but the 65W vs 210W TDP means the GTX560Ti448 uses ~3.25 times the power of a GT 640.
If you had the slots/PCIE bandwidth you would be better off with two GT640's. That said, if I was thinking about 2 I might wait for a GTX660, of some description.
Anyway, as MrS said it's very much down to the PSU.

PS. In the UK a GTX560Ti448 costs ~£190 while a GT640 costs ~£75, so for crunching here the GT640 wins hands down; ~37% better crunching performance per initial outlay and 75% cheaper to run.
____________
FAQ's

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

Waiting sounds about right for you then. And a quality 650 W PSU won't have any problems pushing CPU, GT640 and a GTX680 :)

You could use your 8400 at POEM now. I've tried it today with an 8400GS @ 1.0 GHz: it's doing insane ~33k RAC over there!

MrS
____________
Scanning for our furry friends since Jan 2002

96-98% GPU utilization is high.
I'm sure PCIE3 will outperform PCIE2 but don't know by how much.

It's normal that different cards require different voltages, though task types can vary temps.
____________
FAQ's

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

This project favors faster, mid-range to high end cards.
The GT640 is an entry level card. While it should still work, I wouldn't recommend a GT640 to crunch here (never did, just said it should work).
The GTX 650 Ti 2GB is probably the lowest GPU I would recommend.
____________
FAQ's

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

Your i5 has an 84W TDP and it's pricey. While it won't consume 84W it might use around 60 or 70W crunching CPU projects. The other components would probably use over 30W. Even a GeForce GTX 650 Ti would use around 100W by itself, so your not going to make 100W starting from an i5. Around 200W is doable though, but that's still not a good setup.
Maybe better to read and discuss this in the following thread,
http://www.gpugrid.net/forum_thread.php?id=3518&nowrap=true#33570
____________
FAQ's

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

As SK said, GT640 DDR3 was never recommended for GPU-Grid due to these reasons:
- It's memory bandwidth starved. Mine runs GPU-Grid at 60 - 70% memory controller utilization, with the memory already OC'ed (not much room). GTX660Ti is already limited by memory bandwidth and runs at ~40% utilization. At Einstein GT640 even runs at 99% memory controller utilization!
- GPU-Grid needs results back fast, hence the bonus for returning WUs early. This makes small cards a bad choice over here.
- For a little more money you can get significantly faster cards.

The new GT640 GDDR5 fares a little better due to more bandwidth available, but compared to GTX650 and higher it's still pretty imbalanced.

And as others have already said: don't buy cards with 1 GB or less for GPU-Grid. That's not sufficient any more even today.

Regarding the idea of a small cruncher to reduce energy cost; I don't think it's a particularly good one due to the following reasons:

- every computer has a certain idle power draw, no matter if you use it or not. For a modern "econo-box PC" that's 30 - 40 W. Which you always have to pay just for the system to be there, no matter how fast or slow it is. If you talk about 100 W overall power consumption that's almost half of the power draw "wasted for nothing"! To summarize: the more of your power budget you actually use for crunching, the more efficient the system becomes.

- PSU efficiency peaks around 40 - 60% load. The smallest high efficiency PSUs are the 400 W FSP Aurum (80+ Gold) and 450 W Antec Earthwatts (80+ Platinum). But running these at 100 W load (25%) misses their sweet spot.. it's not a catastrophe for the small Aurum (~4% loss), but any bigger PSUs really start to suffer, which reduces system efficiency.

- Starting from low-end GPUs there's a range where every x1% more purchase price gets you x2% more performance - with x2 being larger than x1. GT640 is borderline in this regard:
GT640 | 691 GFlops | 28.5 GB/s | 65€
GTX650Ti Boost | 1505 GFlops | 144.2 GB/s | 115€

The comparison doesn't look too bad for theoretical GFlops: GTX650Ti gets you 2.2 times the raw performance for 1.8 times the price. But GT640 can't make proper use of its raw horse power due to being bandwidth limited.. which the 5 times higher value of the GTX650Ti Boost hints at (although this card has a bit more bandwidth than it needs - GTX660Ti has just as much, but wants to sustain 2460 GFlops from that). I'd estimate the overall performance difference between GT640 and GTX650Ti boost in the range of 2.5 - 3.0.

- Faster GPUs will have better resale / reuse value.

What I'd do in your case: pimp your primary rig a bit and you may be able to make it fit for 24/7 crunching, or just a bit more than it's doing now (depending on how much you want to invest). By running the faster rig a bit more often you could get the same throughput but wouldn't have to buy entirely new parts. A few points to consider:

- What PSU do you currently have? Exchanging it for a 80+ Gold model in the 400 - 500 W range (e.g. that 400 W FSP Aurum) could quickly pay for itself by reducing energy cost, depending on what you're currently using.

- Is you CPU heavily overclocked, since you're using a "K"? If so energy efficiency obviously suffers, a lot. In this case consider taking it back a few steps. My i7 3770K is running at 4.10 GHz at 1.03 V - that's pretty efficient and for me OK to run 24/7. And I wouldn't feel a difference compared to e.g. 4.5 GHz at 1.2+ V anyway, except for the additional heat, noise and electricity bill.
If you don't OC: you could lower your CPU voltage significantly below stock (as I have at 4.1 GHz), which should save ~20 W over the stock configuration.

- You could adjust your GTX570's clock speed and voltage down a bit. Fermi doesn't have much room for improvements here, though. Increasing fan speeds to lower temperatures could also gain you a few W - if the noise was OK (probably not).

- You could exchange your GPU for a medium sized Kepler. Performance would stay about the same, while power consumption may be lowered by up to 100 W under load. Running 24/7 this would save you ~200 €/year, which means a GTX660Ti would pay for itself within one year. And can be power-tuned well by reducing the power target.. down to about 100 - 110 W for the card under load (stock: 130 W, it's power target)

- If you're crunching on the CPU you could stop it altogether to get power consumption under control. This depends on your project choices, of course. For soem credits count (-> GPU), while for others WCG badges (CPU time) are holy.

MrS
____________
Scanning for our furry friends since Jan 2002