geforce 1080 vs 1070

[beer]

hi
I am looking for upgrading my geforce 970 since it is a bit old.
I am in doubt which chip I should focus on (eg 1080 or 1070).

I am using my computer for a few games as well but they dont max out either chips so the most important is folding@home performance. It should be mention that I pay around 0.3 dollar pr kW

[foldy]

I would recommend a GTX 1070 because it is only 10% slower than GTX 1080 in folding but price is 450$ instead of 700$.
Compared to GTX 970 the GTX 1070 will be twice as fast, 2xPPD at same 200W power.
Do not use the founders edition because custom designs keep the GPU cooler which allows max boost clock and are silent.

[FldngForGrandparents]

Might want to look at the AMD RX 480 coming out in a couple weeks as well. Two RX 480's look to be faster than a GTX 1070, will use the same power, and leave you with $50 or so for beer.

[7im]

Might want to look at the AMD RX 480 coming out in a couple weeks as well. Two RX 480's look to be faster than a GTX 1070, will use the same power, and leave you with $50 or so for beer.

But not the same heat output. Save the $50 plus a little for the AC bill in the summer.

And wait for some real world PPD numbers.

[7im]

This forum support people with problems associated with FAH. If you want information about overclocking brand new hardware, you'll have better luck on one of the overclocking forums or even from a simple google.

[foldy]

No problem running a gtx 1070 together with a gtx 1080. Both share the same driver.
For optimal performance your PC should provide two pcie 3.0 8x speed slots and your power supply must be strong enough and add some case cooling.

[1337LutZ]

I would take the GTX 1070, I took it since I would be gaming on 1440p which is fine with that card.

Overclocked I pull about 780K PPD on average with spiked up to 980K ppd.

[Foxbat]

Interesting observation: I'm not too sure how critical it is to run these GTX cards with "late model CPU" and "pcie 3.0 8x" motherboards. I'm using the Ubuntu 14.04 LTS with the latest July NVIDIA drivers on my two Folding towers, a Mac Pro 1,1 and a Dell Precision T7400. The Dell has my EVGA GTX 1070 SuperClocked and (after some clever power harvesting) my EVGA 960. According to the NVIDIA settings program, both cards are using between 1 and 2 percent of the PCI-e 2.0 slots's bandwidth. The Mac Pro 1,1 has PCI-e 1.0, and over there the Zotac GTX 970 uses 2% of the bandwidth of the PCI-e 1.0 x16 slot but the Zotac GTX 750 Ti plugged into a PCI-e 1.0 x4 slot is using 25% of the bandwidth. That's all during Folding with Core18 and Core21 Clients.

Just a quick look with FAHControl on my iMac shows the follow PPD estimates:
Mac Pro w/2 2.66 GHz Dual-core Xeons - 352,956 PPD
* CPU (2 CPUs): PRCG 9036 (186, 8, 15) - 2,628 PPD
* GPU0 (Core x21 on GTX 970): PRCG 11706 (0,195, 93) - 303,463 PPD
* CPU1 (Core x18 on GTX 750 Ti): PRCG 9162 (81, 0, 446) - 46,865 PPD

Dell Precision T7400 w/2 3.0 GHz Quad-core Xeons - 858,239 PPD
* CPU (6 CPUs): PRCG 9033 (42,7,12) - 14,511 PPD
* GPU0 (Core x21 on GTX 1070): PRCG 13500 (0, 279, 2) - 685,739 PPD
* GPU1 (Core x21 on GTX 960): PRCG 13201 (23, 0, 190) - 157,989 PPD

My Mac Pro I've had Folding since April 2007, first with the beta multicore Client, and after Apple dropped support, I moved it over to 64-bit Ubuntu 14.04 LTS and added a GPU. It's still cranking out respectable numbers after all these years as the GPUs are the stars of Folding. Its biggest issue is the power for the GPU cards as there are only two 6-pin PCI-e VGA connectors available.

So, I maintain that you don't need the latest and greatest CPUs or Motherboard tech, just a solid power supply with plenty of 6-pin and 8-pin PCI-e VGA connectors and Maxwell and (now) Pascal boards. Just get out there and Fold!

[1337LutZ]

I would take the GTX 1070, I took it since I would be gaming on 1440p which is fine with that card.

Overclocked I pull about 780K PPD on average with spiked up to 980K ppd.

Goodness!!! Have you had any stability or temperature problems with overclocking the 1070???
Has anyone tried overclocking the 1080?

I only wish everyone had a late model CPU and a 1080! But, wishing never made it so. All we can do is what we
can do.

Are we any closer to solving the folding problem than we were when I started F@H about a decade or so ago? Will
we ever have enough computing power to bring this complex problem to solution? Perhaps "quantum computing" will
advance a solution, but that's a decade or two away from common use.

I installed a Kraken G10 and a Kraken X31 AIO cooler on it to give it a constant 58degrees whilst being quite silent with these settings:

108% power limit
100% voltage boost (allowing for 1.09 volts)
+150mhz core
+700mhz memory

totalling a 2114mhz core clock whilst folding and 8972mhz memory clocks.

Core clock increased the PPD by a fair bit, but that 700mh memory clock added another 30-50k PPD! which was new to me, since usually memory does nothing.

[foldy]

@Foxbat: Interesting because on Windows the MSI Afterburner shows 50% bus usage with a gtx 970 on pcie 2.0 x8 which matches your pcie 1.0 x16 but you only get 2% bus usage.
Is your linux nvidia settings showing the right values? Or is bandwidth usage different in Windows or Linux?

If you want to do an experiment then switch the gtx 1070 and gtx 970 to see if the PPD changes on one of them.
Expected behavior in theory is gtx 970 gets 10% more PPD and gtx 1070 gets 50% lower PPD.

[Foxbat]

Yeah, good point. I guess I'm trusting NVIDIA's software to provide good feedback. I just took it to mean that once the atoms are loaded into the GPU's memory, there isn't a lot of traffic back and forth. I'm probably very, very wrong in that assumption!

I just got the GTX 1070 up and running so I'm going to let everything sit for now. I'm not sure if the power requirements for the EVGA GTX 1070 SC can be satisfied with the stock P/S in the Mac Pro 1,1.

[rwh202]

Yeah, good point. I guess I'm trusting NVIDIA's software to provide good feedback. I just took it to mean that once the atoms are loaded into the GPU's memory, there isn't a lot of traffic back and forth. I'm probably very, very wrong in that assumption!

Yeah, this makes more sense than the huge bandwidth usage often quoted on windows - I can't believe that it it's really using 2 GB/s continuously (50% of 8x gen2). I'm seeing 1-2 % of gen2 4x on linux for a 980. I suspect that the 'usage' shown in windows might be similar to the cpu usage used by the drivers - it's just polling, not actually doing anything useful most of the time. I'm going to try a 1080 on a gen1 1x riser, and I'm not expecting to see much of a slowdown, but will report back after experimenting.

[Sven]

I'm going to try a 1080 on a gen1 1x riser, and I'm not expecting to see much of a slowdown, but will report back after experimenting.

That's going to be an interesting number.

[rwh202]

I'm going to try a 1080 on a gen1 1x riser, and I'm not expecting to see much of a slowdown, but will report back after experimenting.

That's going to be an interesting number.

I've had a quick play and gathered one datapoint. I wasn't too keen on leaving a 1080 jerry rigged all over the lounge floor, so will try and put something together in a spare system later.

GTX 1080 FE (stock), linux mint 17.1, driver 367.27, core 21 v0.0.17, p9211 (r59)

PCIe 2 16x tpf: 4:46 = 500 kPPD (1% bus utilisation, 98% GPU utilisation)
PCIe 2 1x tpf: 5:00 = 466 kPPD (13% bus utilisation, 93% GPU utilisation)

So in this instance, there is a 5% difference in raw performance when bandwidth is reduced by 93%. Different projects/cores/drivers might be affected to a different extent, but I certainly wouldn't lose any sleep over using a 4x slot for even today's high end GPUs. However, what may still be a factor is how those PCIe lanes are connected to the processor - in my tests, both examples were using the CPU's dedicated graphics lanes. Going via the PCH and any other switches etc. may introduce more latency to main memory or whatever resources the cores are communicating with. Again, hoping to dig out an old system with a variety of slots and experiment a bit more.

[Nathan_P]

When I tested a couple of months ago on PCIe 3 x16, PCIe 1 x16 and a couple of combinations in between I saw a 2-6% PPD drop on a 970 and a 980ti