Anyone with cMP and Titan-Xp or 1080Ti and power issues?

[Rominator]

I don't think it's slot draw.

I can solve problem by moving the 6 pin to another source.

But it is weird that issue is caused by very fast spike that doesn't show in HW Monitor.

[mysticalos]

It's probably a power management issue with how fast the power supply can ramp up current. the cards are very conservative with power not under load. it sounds like power droop to be honest. droop is something OCers have to deal with all the time to ensure sudden ramp up doesn't do exactly this.

[Squishy Tia]

I don't think it's slot draw.

I can solve problem by moving the 6 pin to another source.

But it is weird that issue is caused by very fast spike that doesn't show in HW Monitor.

The reason I'm betting on slot (PCIe bus) draw is that unlike many PC mobos, Apple's Mac Pro mobos do not have a PEG connector to supplement power to the PCIe bus. This perfectly explains why the Pascal cards shut the machine down outright when another PCIe card is in use and when CUDA-Z ramps things up in Heavy Mode. Oddly enough, this actually contradicts my concerns regarding using SATA power to supplement the GPU power requirements because that's exactly what a PEG connector does on PC mobos. The main concern then really, is any device on the SATA cable in addition to the GPU's PCIe connector - any power draw spike (such as CUDA-Z) could potentially fry any device connected in the daisy chain on the same SATA cable, and since Apple uses a single SATA cable with dual headers, you see the problem.

The only (relatively) safe configuration for Pascal in a Mac Pro is muxing the two mini-PCIe 6-pin into a single 8-pin via the appropriate adapter and using the SATA cable for the remaining 6-pin connector on the GPU and not using any additional PCIe cards, as the MPs don't have a PEG connector to supplement bus power to the slots.

You can have both a power draw spike and a power "droop" at the same time, ironically. The spike in amperage can actually lower the voltage enough to trigger a fault and shut the machine down even if the total wattage is within spec. Apple's PCIe power path on both the slots and its two 6-pin connectors is designed for steady loads, not spikey loads. Those mobo traces aren't designed for high demand loads like full-on shielded and insulated thick gauge cables are.

Bottom line is this: If the Pascal card has two 8-pin connectors, there is no safe way to run it in a Mac Pro outside of using an external PSU such as one of those that goes into the optical bay and connects to a real plug through the PCIe slot exposure in the rear via a specially designed cable. And the only workable way to run any other Pascal GPU that has a mix of 6 and 8 pin connectors (one of each) is mini-PCIe 6-pin x2 -> PCIe 8-pin x1 adapter coupled with SATA -> 6-pin PCIe adapter for the remaining 6-pin connector, and you can forget about using any additional PCIe cards.

Mac Pro PSU can handle the spikes. But the motherboard cannot.

[netkas]

Doubt its slot power, check this article : http://www.tomshardware.com/reviews/nvidia-geforce-gtx-1080-pascal,4572-10.html

clearly power spikes on 8-pin port, pcie slot isn't getting to 75w even in furmark

quick update on 1080 egpu power issues.
I use 550w psu for it, with 456Watts on +12v.

old config was - barrel plug goes to 12v on atx port (the one you plug into mobo)
pcie power is from psu's 6+2 power plug.
launch unigine superposition - system crashes.

now I try to more evenly distribute the load across +12 v lines (seems the psu doesn't have common +12v line)
using 2*6pin -> 8-pin , one 6-pin is directly from psu, one 6-pin is combined from two molexes from different +12 lines.

seems to work so far.

[Squishy Tia]

The eGPU situation would also affect PCs that have PSUs with multiple rails that aren't adequately configured in that case. My guess would be a single rail PSU is ideal for the Pascal GPUs, preferrably one with at least ~90A peak rating on the +12v line. Do you happen to know what the max amperage rating is on your eGPU's rails?

[netkas]

http://www.tomshardware.com/reviews/nvidia-geforce-gtx-1080-ti,4972-6.html
at max 55 w from pcie slot for 1080ti
but about 160-200 watts from two power slots.
I would assume the card suck more from 8-pin
so combining two mp's 6-pin into 8-pin and powering card's 6-pin from two sata ports might do it.

[Squishy Tia]

There is one possibility, a trick I used while still on my original Mac Pro. Normally Apple's fans don't ramp up until the temps really exceed a certain point. I found out the hard way when my then new Radeon 5870 Mac Edition would begin sounding like a dentist's drill even at only moderate load while the Mac Pro's fans stayed silent at their default 650 RPMs. So what I did was use SMCFanControl to create a gaming profile where my PCIe fans would spin at 1100 RPM. Much faster, but still essentially whisper quiet due to the Mac Pro's design. Perhaps MP users could use a similar profile setup to keep the GPU cooled at all times, thus creating a situation where thermal throttling or heat related spikes don't kick in at all. It's worth a shot. That's one of the few advantages the MP design really had going for it - the user could force the machine to keep cool and still maintain a whisper quiet profile.

Now, the tomshardware.com test was done on a FE card, which uses a shit blower, so anyone that has a better cooler on their GPU may end up with less of an issue in this regard. Being a 300w TDP card though, it really does look like the load on the 8-pin connector is maxed out, indiicating nVidia did a poor job with power distribution design. It's interesting to see the GPU thermally throttle so easily though, something I wouldn't have expected from a Pascal GPU. And we both know how temperature can affect power draw. Perhaps the Mac Pro's unique design and airflow setup combined with SMCFanControl can create a more stable environment than your typical PC case with off-axis intake/exhaust. When the GPU gets cold air vs. moderately warm air, it can make a massive difference.

I'm actually kind of glad I have airflow overkill in my Cosmos II case for this hackwell pro. Might make a mountain of difference w/ a 1080Ti.

[Rominator]

Nothing to do with temp.

I can start cool machine with cool 1080Ti, run CUDA-Z and BANG it's off. Card hasn't even warmed up.

It is 100% fast current spike on the power plugs that I can fix by moving the source of the plugs.

[Rominator]

So...we got an eGPU rom for the Titan Xp that enables OS X and Windows on nMP without a SINGLE SOLITARY SYSTEM FILE MOD.

Plug it in to a TB chassis and fire up the 6,1 (or 2014 Mini).

Ran some tests with 1080, then plopped the Xp into Netstor TB2 chassis which has never had ANY trouble with Maxwell cards, up to and including Titan X.

I set 3DMark to run "Timespy" and left the room. When I came back the nMP was running it's fan like a whirling dervish but the screen and Netstor were black.

Unplugged and rebooted. This time I stuck around to watch. As soon as Timespy got to running it shut off like a light again. Netstor has Dual 8 Pin inside. But apparently not enough.

Good news, I plugged a Razor Core in and it got through Timespy. So Razer took the 8 pins a little more seriously. Akitio Node will be next.

But by definition, the Pascal Xp is drawing more current than Maxwell Titan-X. I'll post the 3D Marks in the eGPU section. Guess I could try in 5,1 with X5680 too. Anyone interested?

There is some discussion at MR about power limits of 4,1/5,1. Somewhere buried a couple years back I did some serious tests and posted them there. 120 Watts is the deep end for those onboard PCIE cables. Anything more and you get the CLICK and darkness. The first Titan and Titan Black could hit that. Weird thing about these Pascal cards is you never see those numbers. I've never seen more than 80 Watts then CLICK and I'm staring at a blank screen. So the spike is much faster than the old Kepler cards would do.

[karaluchy]

Please continue the investigation, I want to buy GTX 1080ti from you for professional (cuda&openCL) and gaming(bootcamp) purposes in my cMP 2x X5690.

[Squishy Tia]

The MP's two mini-PCIe connectors top out at 120w total?? That's not even close to the spec of 75w per 6-pin connector. If that's the case, Apple seriously fucked up using shitty traces to provide the power through the mobo.

[lowrider]

All the talk is about the PS.  What about the MB traces when they card draws all that current?  The main reason I went for the GTX 1080, was the one 8 pin connector and the lower TDP than my previous GTX 780.   

Lou

[Rominator]

The MP's two mini-PCIe connectors top out at 120w total?? That's not even close to the spec of 75w per 6-pin connector. If that's the case, Apple seriously fucked up using shitty traces to provide the power through the mobo.

120 watts is highest peak I've seen that hasn't lead to shut down on a single cord

I've never seen both in this area at same time as most cards draw more from one side than the other

[karaluchy]

So to sum it up, the cmp mobo is too old and shitty to power the 1080ti or titan Xp? 

[Rominator]

So to sum it up, the cmp mobo is too old and shitty to power the 1080ti or titan Xp? 

Well, that's the "glass half empty" version.

Or you can say "Gosh, who needs 6 SATA drives, I'll grab power from one of those and put a 2017 GPU in my Chesesegrater" and be the "Glass half full with a silver lining" type guy/gal.