Help with setup of i5 4590 plus nvidia GTX960

I won't pretend to know optimum, but for people interested in economy in both purchase and operating cost, I think that a quite inexpensive Haswell dual-core CPU hosting one or two GTX 750Ti cards is something worth setting as a comparison reference to try to decide why a specific configuration under consideration is better for a particular user's interests.

Of course, neither that CPU nor that GPU will give a romping stomping system competing up near the top of the RAC rankings, but if one is careful on spending on the other components--or better yet is in a position to re-use existing case/power supply/disk hardware, then the purchase price and operating cost for such a choice can be rather far below most of the leading configurations.

My biggest concern with actively suggesting either of these parts (as distinct from proposing them as a reference comparison) is that they are at the point of likely being superseded. In the normal Intel progression, most likely some Skylake offering may soon offer somewhat better speed/power/price tradeoff than Haswell, or quite well may already. The Nvidia sequence has taken an oddly long pause at the 28 nm TSMC process "node", so the very next offering, slated to be Pascal series parts on the TSMC 14 nm finFET "node" may offer a considerable improvement in performance at a given power consumption point.

Skylake is here and now, though somewhat early days. Pascal is still in preshipment hints, winks, nods, and sightings of alleged prototype sample chips in India customs paperwork. I've seen no credible claim as to whether the first Pascal part offered will be the "big chip" probably meant to displace GTX970/980 or a "small chip" meant to display the GTX750/750Ti. In the previous generation there was considerable delay between these, even though the supplier presumably had plenty of fab capacity. This time TSMC finFET capacity and yields may both influence this choice. Still, some Pascal flavor will surely ship in 2016, or Nvidia has had a big setback.

In counting the cost, one must not neglect the power supply, though sizing it properly is a bit vexing. An advantage of a single 750Ti system on a dual-core Haswell is that very likely any supply you might already have may be sufficient. Even a dual 750 system does not need a very capable supply at all (and if I were selecting a motherboard for such a system today, I'd try to find a economical one which had two side PCIe slots, just to protect the possibility of going double later).

I used to be somewhat skeptical of multiple GPU systems, believing from some data on other people's systems that the extra GPUs were so slowed below maximum output by sharing host support as to deplete the economies considerably. But the current Einstein CUDA55 Parkes PMPS application seems to "share" very well indeed (at least if you don't load up the CPUs with BOINC CPU tasks--I run zero), and from both a purchase economy and an operating cost point of view, sharing the host cost over two cards is a very big advantage.

Of course, this may all change with another application release, and all tradeoffs may be different for work from another project, gaming, or just a different application on this project.

This thread about getting started with GPUs, is worth referencing, probably worthy of stickification.

That was my design decision 3 years ago. I did say that I would most likely go for two GPUs now.

Back then, the GPU apps here were very hungry for PCIe bandwidth. Two or more GPUs would be sharing that bandwidth so the potential was there to have an adverse effect on the crunch times for each GPU. There have been changes to the apps that reduce the bandwidth requirements. Its not a problem any more.

There have been improvements in power consumption for both CPUs and GPUs. You can't really rely on what worked well three years ago. I would have no qualms in agreeing with a design just like archae86 mentioned - a Pentium dual core with two 750Tis.

The motherboard you choose is probably a more important decision than the CPU. With the latest apps now much less reliant on high PCIe bandwidth, you don't need the very expensive boards that allowed the full complement of PCIe lanes (x16) on each slot. I'm sure x8/x8 would give you full performance on both cards. I've noticed some quite cheap boards that are shown as x16/x4 in the specs. I don't know if they could be configured as x8/x8 when two similar cards are installed. Maybe the x4 slot might still be OK for the performance of the second card. These are the sorts of questions you need to address. Maybe archae86 might comment on the boards he uses.

In the future, particularly at Einstein, the importance of CPUs is going to continue to diminish. Eventually, I would expect most (if not all) science runs here to have a high performance GPU app available. It might take a year or so but it should figure in your planning for a crunching box. If you're serious about crunching in the future, save on the CPU cost so that you can spend on a GPU upgrade down the track a bit. That's where you will get most benefit from your spend.

A Pentium dual core is perfectly adequate to service two 750Tis each crunching at least 2x. I would expect that you could crunch CPU tasks on at least one of the cores, maybe even both with little effect on GPU crunch time. It's actually quite interesting to experiment and find out. You would probably get the best credit output to power consumption ratio by disabling CPU crunching completely. So, it really depends on what is most important to you - saving electricity or helping to find a new Gamma-ray Pulsar ;-).

(Pssst - don't look now but I think you've really caught the crunching bug!!) :-).

My most recent build in early 2014 used an ASROCK Z87 EXTREME3 motherboard to host a dual-core Haswell. At the time of purchase the price was $115 US, so not a true economy board but not remotely high-end. The full-length PCIe slots on that one configure as x8/x8 with one GTX970 and one GTX750Ti plugged in. I no longer recall what considerations drove me to that board rather than one of the sub $100 boards at the time.

I recently reduced all three of my main crunching machines from running just one CPU job to running zero, expecting to get a power efficiency of computing gain at some modest loss in total Einstein production.

I did indeed get power efficiency improvement from each system, but on the dual core Haswell system in question I was quite surprised to see sufficient output increase from the GPUs to actually get a noticeable net Einstein production increase.

One possible interpretation of that result is that my dual-core Haswell on that motherboard is not quite fully keeping up with the needs of the combination of a 970 and a 750Ti, though I think it is doing pretty well.

Experimentation is key, as there are enough variations in system configurations and applications to make predicting your results from others' experience (or even your own experience) surprisingly iffy.

There are some very low-priced Skylake CPUs sold under the Pentium brand these days. I recently reviewed a list of the capabilities disabled by the manufacturer on these chips as compared to i3/i5/i7 chips sold from the same die, and got the tentative answer that for my purposes they would probably be entirely adequate. "Pentium", "Core", i3, i5, i7, and even Xeon are just brands, and as with, say, Ford/Mercury/Lincoln, the names are not a durably reliable indicator of the detailed characteristics of the products sold under that name in any given year. Anyone telling you what a Skylake Pentium is like based on their memory of ancient products sold under that name is speaking through their hat.