Mac OS X - Multiple GPU Mining? Possible?

[catfish]

My Mac Pro had a silly-money 4870 in it - supplied as an upgrade by Apple. It took two 6-pin power adaptors.

Since I had a 6870 lying around, which wasn't particularly spectacular as a miner in my Linux farm, and I've started using two-to-one 6-pin Y-cables for neatness, I decided to chuck it into my Mac Pro (which has adequate power headroom and a clever thermal zone system).

OK, I had to do some hacking to put an EFI section into the Radeon BIOS so the card would display the boot screen. But other than that... the 6870 works just perfectly in my Mac Pro - and an upgrade from 512 MB of video RAM to 1 GB makes the Mac GUI a LOT more responsive (I have 30 inch screens and have a LOT of windows open at the same time...)

The big question - OpenCL.

Well, I assume I now have two GPUs - the 4870 and the 6870. The 6870 is connected to the main 30 inch screen and I'm using it now - it works. The 4870 is connected to another screen and it works. I can run two instances of phoenix, using DEVICE=0 and DEVICE=1.

Using DEVICE=0, which is reported as 'ATI Radeon BA Prototype' for some reason (probably another netkas joke - I used his BIOS hack), I'm getting a steady 145 MH/sec, using poclbm and the flags I'd *normally* use for mining. Remember that this card is also powering my main desktop, the 30" panel with 6 'Spaces' - each with hundreds of windows open.

Using DEVICE=1, which is reported as 'Radeon HD 4870' as expected, python crashes with a segfault, traceable back to ATIRadeonX2000.kext.

I'm absolutely SURE that I must have OpenCL working on the hacked 6870 - there's no way a 4870 would be delivering 145 MH/s, the best I got in OS X was 90 MH/s under similar conditions.

However I've been running five instances of phoenix on five separate cards under Linux. And I haven't changed the software since running phoenix (and poclbm, with the same settings) on the Mac Pro when it only had the 4870 installed (getting 80-90 MH/s).

The big issue here is that pyopencl on OS X links to the Darwin kexts and not to the Linux AMD SDK, obviously. Perhaps the limitation is there.

Anyone else messing about with their Mac Pros, or more commonly, the big Macbook Pros with dual GPUs? Is this a limitation of Snow Leopard OpenCL frameworks, or does the Mac fork of pyopencl need tweaking?

(by the way, if you've got a Mac Pro, and are considering paying £370 for Apple's graphics card upgrade, do yourself a favour and buy one of these PC cards... the 6870 isn't anywhere as fast as a 5870 but it's plug and play - I didn't even need a PC to flash the thing...)

[1503230066]

1503230066

[1503230066]

1503230066

[1503230066]

1503230066

[1503230066]

1503230066

[rethaw]

Perhaps you have resolved this, but I can tell you Apple's OpenCL implementation is for from perfect.

I have had the best luck using DiabloMiner on Mac, you may give that a try and see how it goes.

[catfish]

Nope, after seeing the light, I'm trying to get cgminer compiled and working perfectly on OS X 10.6.8.

I've been through the CCFLAGS dance already having compiled the ArtForz cpuminer (for Litecoin mining) on this particular Mac (Mac Pro 3,1 - 8 cores, Xeons, still reasonably nippy). However I've got past the issue that is reported on these forums (i.e. compiler can't produce executable output) and am still stuck with a complaint about libcurl.

It simply could be that the libcurl that I've installed via ports isn't extensive enough and I'll need to download source and compile (and hope that the authors took OS X into account), it may be possible to hack a tweak into the configure script.

However I think cgminer will be the way forward for OS X.

I've also got a Hackintosh coming on line that is likely to have three highly overclockable Sapphire 5770s (small beer, I know, but not bad inside a case with a hot i5-2500 CPU). It's just a case of using up my remaining hardware now - additional investment is out of the question, until I have capital for FPGAs


Rethaw - regarding Apple's OpenCL implementation, it is poor on Snow Leopard, outstandingly good on a release candidate of Lion, and average-to-good in retail Lion. For ultimate performance, the trick is to get the kexts from that particular Lion release candidate (they may even work in teh ice kitteh too).

But, as always, the main reasons Macs don't make ideal GPU mining boxes are that 1. Mac high-end AMD GPUs are somewhat expensive, and 2. Profitability is all about overclocking, and OS X doesn't have API calls for overclocking the cards (or at least I've never seen any).

However, they are energy efficient, quiet for their heat output and reliable. All my dedicated miners are a self-optimised distro of Linux with homebrew monitoring tools, of course... but it wouldn't hurt for my workstation Macs to add a few hashes to the farm, would it now?