Differences between Intel iGPU work units and Nvidia / AMD

Because your computers are hidden, all of us mere mortals haven't really got a clue about what you have and what you might be able to "wring out of them" :-).   You need to specify, in a bit of detail, the CPU model, RAM, GPU, etc, of particular machines you might be commenting on.  There have been a number of 'generations' of Intel iGPUs and there have been many issues about driver versions that would allow particular models to return valid results.  There are long running threads (like this one) about this that you could peruse.

If you want to know what type of tasks are available for particular GPU types, just go to your account dashboard and look down in the bottom right hand corner for the 'applications' link.  If you open that page you will see that the only work available for Intel GPUs at the moment comes from the Arecibo binary radio pulsar search.  The tasks in this search are deliberately 'small' as they are designed for 'mobile' type devices and Intel GPUs.  In the past, much larger tasks were available for discrete GPUs.  It's unlikely they will return because the data supply can't keep up with the demand.

There is a much greater supply of work for discrete GPUs from the gamma-ray pulsar binary search (FGRPB1G).  These are 'large' tasks, suited to modern mid-range to high end cards.  The app uses OpenCL (and not CUDA) which is why older NVIDIA cards struggle in the way you describe for a 560Ti.  At some point in the future, there may be a CUDA app which should improve this.

That's quite OK.  I have pretty much the same attitude to privacy matters.  The older I get the more I tend to cherish what might remain of mine and despair at how willingly some much younger people seem to abandon theirs :-).  For details about hosts I own, I usually provide a link to a hostID if I want to share the details of that host.  That way people can see both the hardware/OS details and the crunching performance.

I've never tried to run tasks on the iGPU because I run Linux and there were no drivers available at the time.  I don't know if that has changed in more recent times.  I'm much more involved with discrete GPUs.  From what I remember by reading what others have posted, performance will suffer with iGPUs if you also load all CPU cores.

With discrete GPUs, the default is to require a 'free' CPU core to support each GPU task - for both AMD and NVIDIA.  You can see for yourself if this is really needed by looking at some hosts with particular GPUs of interest, ie., in the top hosts list.  All NVIDIA GPUs I've ever looked at have CPU times which are almost as large as the elapsed time.  In other words, a full CPU core is being 'used' for pretty much the entire duration of each GPU task.  So, the default really makes sense for NVIDIA GPUs.  It's quite a different story for AMD GPUs.  The CPU time used is very much smaller than the elapsed time.  I've found from experience that for the GPUs I use, I don't get any significant degradation in performance even from running as many as 4 GPU tasks over 2 GPUs with just a single CPU core (from a Pentium dual core processor) for support.  The GPU model is R7 370 (for both) and the CPU component is only ~140s out of a ~2050s elapsed time.  I use the app_config.xml mechanism to remove the default allocation of 1 CPU core per GPU task.  Running 4 GPU tasks actually 'reserves' zero CPU cores.  I reserve a single core by setting computing preferences to 'allow' BOINC to only use 50% of the cores.

That comment was about discrete GPUs.  I don't believe there have ever been 'larger' tasks for iGPUs.  There would be a lot of 'back end' overhead in maintaining different task sizes for 'mobile' devices as opposed to iGPUs so I wouldn't expect any change in the current situation.

I specifically said that older NVIDIA cards struggle.  The top hosts list shows that recent generation high end NVIDIA cards do very nicely, thank you, even though you still need to use the default 1 CPU core per GPU task for support.  If you have the money for a 1080Ti, don't let my comments discourage you in any way.  The evidence shows it will crunch very well if you provide the support.

My philosophy is that I object to paying an unfair premium just because a company decides it can get away with it.  The classic example is the high end Intel processors which have suddenly become a lot cheaper now that there is viable competition.  In this country, there is an 'Australia Tax' as well where components are significantly more expensive than what you would pay in the US, for example.

So I tend to take advantage of significant discounts available in older models when the new replacements are first launched, particularly if it appears likely that the replacement will have only minor potential for improved performance anyway.  This seems to work quite well for AMD but not so much for NVIDIA.  The latter always seem to be disproportionately expensive here so I end up buying AMD most of the time.  That does mean higher power costs but that has changed quite a bit with the RX 400 series and should be even better for the next series.

OK, I'm wondering if I haven't explained things properly.  Since the current FGRP GPU app is a completely different beast from what went before when the BRP4G/BRP6 searches were in progress on discrete GPUs, I neglected to point out the big change in CPU support needed.  At the time of BRP searches, NVIDIA apps were based on CUDA and the default CPU 'support' estimate was set at 0.2 CPUs per GPU task.  There was essentially little or no penalty if you ran 2 or 3 GPU tasks on mid range GPUs without 'reserving' any CPU cores.  If you weren't modifying the default with the app_config.xml mechanism, you would have needed to be running 5 concurrent GPU tasks before any automatic 'reservation' of a CPU core by BOINC would have kicked in.

With the new OpenCL based FGRP GPU app, the default was immediately to reserve 1 CPU core per GPU task and this is what you are obviously alluding to with your comment, "changed recently from .2 to 1".  If you weren't doing anything to modify the default, each new FGRP GPU task would have automatically restricted any CPU task crunching by one core anyway.  It's unlikely you could do any better by further reserving an additional core over and above what BOINC does automatically.  I have GTX650s which don't improve at all, even if no CPU tasks at all are running.  The crunch times are very slow, the screen lags are woeful so I've shut down many of these hosts.  I'd certainly restart them if a better performing CUDA app was available.

A 5870 would probably do quite well here.  Its major problem would be power consumption compared with more modern cards.  Heat issues are usually fairly easy to fix.  The two most likely causes are (i) thermal grease which has deteriorated to become essentially an insulating layer and (ii) a fan with bad/dry bearings that stop it from running at the proper speed.  I have a bunch of 4850s - used at Milkyway until quite recently when they did a server 'upgrade' that prevents them from being recognised.  So they just gather dust now.  They only support OpenCL 1.0 and the minimum here is 1.1 so I've never been able to use them here.  They crunched at Milkyway for about 7 years and most had heat problems at various stages.

If your 5870 is anything like my 4850s, it's a pretty trivial exercise to replace the thermal grease if that's the problem.  I've done that on most of mine and it immediately dropped the temperatures by a large amount (100C+ -> 70-80C).  If it's the fan, it may be possible to re-oil the bearings.  Mine were sealed but I was able to drill through the plastic and inject some oil with a diabetes syringe.  I found it would only last for 6-12 months so in the end I zip tied a decent sized server style fan directly to the heat sink.  That seemed a fairly permanent solution as I never needed to revisit any of the ones with that 'mod' :-).