Folding Forum

I didn't want to hijack veedee's topic but I'm in a similar position, except that I haven't bought a GPU card yet

I've got a low-ish use PC (3GHz Core i5, Linux Mint) and thought about putting a GPU card in it to boost its folding work (the CPU is already busy doing that in its spare time), but it's power supply is only rated at 300W- not enough to run an NVidia 460.

I know nothing about what's available in the GPU market, is there one which will give me a reasonable boost in folding power and fit within the constraints of the PSU?

(Preferably without stealing a core from the CPU, but I'm not dogmatic about that)

Depends on how much spare wattage you have. Start with a GT 640 at 40w or a GTX 650 at 64w TDP, and work your way up from there. 650 Ti is ~110w TDP, 660 is 140w TDP, etc.

I didn't realise they started at as low a power as that… I must have been led astray with all the power bills that people are talking about on here

Time to look up some numbers to see what will fit, thanks

Yes, GTX 650 Ti - 105-110 W TPD, but actual consumption is about 65 W (with core_17; PPD - 30k @ 1100 MHz).

artoar_11 wrote:Yes, GTX 650 Ti - 105-110 W TPD, but actual consumption is about 65 W (with core_17; PPD - 30k @ 1100 MHz).

That sounds promising, it's even within my price bracket!

I see the card is PCIe 3.0, the PC motherboard is only PCIe 2.0, which might slow it down a bit? I don't know how demanding the client is about bus transfer speeds. But it should still be a substantial improvement over the CPU!

I'll check with the PC supplier tomorrow that the PSU won't object, but that looks like the best one to go for, thanks both.

No speed loss there.

The influence of bus speed on FAH depends a lot on the size of the WU. There's no absolute rule: All will work though the amount of degradation will vary from negligible to noticeable.

Project 8900 has 40000 atoms which I'd call "moderate" so a mid-range bus speed is probably a good idea. Although you're not able to select specific projects, those with as few as 500-600 atoms ran nicely on V2.0 at 4x. With ~100x as many atoms you may notice some variations between 4x and 8x and 16x. Pushing onward to v3.0 can probably wait until bigger proteins are ported to the GPU.

My MB is also PCI-Е 2.0, no bottleneck about bus transfer speeds.

I can't find anything in the BIOS to set the bus speeds- there's one x1 slot and one x16 slot, so guess which one it'll go in!

I'll likely have more questions when I get to install it (especially re drivers- I gather the older ones are better?), but they'll be best in the appropriate section I assume?

In BIOS - PCI-E 2.0 x16 (slot). I have installed NV driver - 327.23 WHQL.

Had some problems with GTX650TI, see Advice for choosing the GTX 650 [Add: GTX 650 Ti OC], but power usage and temperatures were fine.

Regarding the folding future with GPU focus on Core17 I would look for a small AMD GPU with reasonable OpenCL performance, see FAHBench (OpenMM 5.1), for example HD 7790 (anandtech folding results) or R7 260X (anandtech folding results = source for graphic below, it shows a range of choices), both Bonaire Chips.

With my second hand HD 7950 the power load running FAHBench Implicit is ~195 watts. Full load SMP (cpu:7) + GPU client ~235 watts. But CPU and GPU are undervolted, in my written testing memo there was a p6098 and p8900 with 16681/90309 PPD, this way 460 PPD/watts.

Be carefull, there is no guarantee for projects availability, rather a trend to more complex calculation, which can only be faced with more powerful hardware.

There's a lot to read there, most of it above my head and anyway it's my bedtime!

I'll study it more closely tomorrow, thanks.

All things being equal between AMD and NV, cost and or PPD, I would chose NV. NV drivers seem to be better than AMD lately, although not perfect, and there is always the road ahead where NV JIT potentially doubles NV performance. NV also works on linux now, AMD not until later, if that matters.

If you do choose an Nvidia GPU, please note that you will require 1 CPU to feed the GPU (FahCore_17) regardless of the OS. However, with AMD GPU, you may not need 1 CPU to feed the GPU (FahCore_17) but it doesn't work in Linux (no ETA as to when that will happen).

PantherX wrote:If you do choose an Nvidia GPU, please note that you will require 1 CPU to feed the GPU (FahCore_17) regardless of the OS.

I thought I'd remembered seeing that, thanks for confirming. Not a problem.

… but it doesn't work in Linux.

But that is, there's no chance of me going back to Windows- decision made for me!

Although I might wait until a backup GPU assignment server is sorted out… I'm a complete newbie with GPUs, not a lot better with Linux, and I'd rather not have too many known potential unknowns all at the same time