Dual GTX Series in Mac Pro 1,1
27 Mar 2016 19:41:41 UTC
Topic 198534
(moderation:
I recently decided to revitalize my 1,1 (upgraded to 2,1 with double 4-core 5365's) Mac Pro rather than selling it on Craig's List, and I updated it to El Capitan, got the latest CUDA drivers, as well as the nVidia driver specific to my version of OSX and build number.
I had a couple of GTX 650's available, so I put them both in and was running FGRP6 tasks at 2 each per card. They were completing ~330 minutes per pair, one card slightly faster than the other, probably due to PCI-E bus limitations (on a 1,1 Pro, with one card in Slot 1 and one in Slot 3, you get 16x Gen 1 and 8x Gen 1 throughput, which is much slower than 2.0 or 3.0).
The 1,1 supports two 6-pin power cables, so I decided to upgrade to two GTX 960 2GB models, each of which needed one six-pin cable. They arrived today, and I opened and installed one for the first test.
The results are disappointing; the card is crunching 2 units approximately 50% faster than a vanilla GTX 650. This is about the performance I would expect from a GTX 950, based on the compute capability of the cards and on my own results from using a 950 in Windows 8. As a test, I pulled the 950 from the Windows 8 machine and put it in the Pro, and the 950 was very slightly outperforming the 960, despite the 950 being in the 8x slot. CPU and RAM utilization are normal; with each CPU under 10% and plenty of RAM available. I am not running any CPU tasks, leaving 8 cores available.
Does anyone know if this is normal behavior? I've seen the thread about the memory downclocking of some Maxwell cards, and I'm wondering if it is related to this (though the 950 is also a Maxwell), or if I have a dud card, or am missing something else. Is there an app I can use to view / tweak the card's performance in El Capitan? I normally use CPU-ID's Hardware Monitor and nVidia Inspector in Windows.
In the meantime, I'm going to let the 960 crunch a while today (and have bumped it up to 3x simultaneous, which appears will have a slight overall improvement after accounting for slower completion) to make sure I get valid results, and will then put it into the Windows machine that normally runs the 950 to see how it compares in Windows.
Dual GTX Series in Mac Pro 1,1
)
I have the 960 in the Win 10 machine now. As I saw in the Mac Pro, it seemed to be crunching 2 units at just about exactly the same speed as the 950 on FGRP6 beta.
I set it to P2 in nVidia Inspector and bumped up the memory clock as described in the sticky thread about Maxwell performance, and it appears to be going about 10% faster now based on percentage completed per second.
I then added the 950 back in; with my setup this causes both cards to be running at 8x on the PCI-E bus. When the 960 wasn't unlocked, it ran at just about exactly the speed of the 950. Unlocking it with nVidia inspector allowed a performance increase of 10%. And then 950 doesn't have a P2 state to unlock -- you can only do basic overclocking.
Well, I've come full circle.
)
Well, I've come full circle. I've found a benchmarking site that indicates the 960 is only about a 10% boost over a 950, which fits my observations. Apparently my other sources were not quite as reliable.
You have just discovered a
)
You have just discovered a truth I found out the hard way.. GTX 950s and 960s are actually outclassed by GTX 770s and 780s for power / compute performance, at least I know this for the Mac Pro side.
I had a 960 briefly and ran it against an EVGA GTX 770 SuperClocked with ACX cooler, Einstein reported right off the bat that the 770 had a higher GFLOPS value in the Event Log. This is under El Capitan 10.11.3 and the latest CUDA / web drivers, I think CUDA version 7.5.25 (now 26.) This was in a 5,1 with 24 GB RAM though.
Even GPUBoss.com says the stock 770 has 40% better FP performance than the stock 960..
I think the sweet spot for most bang for the buck would be a 780, but since you can find 770's going for less than $200 refurbished, why bother..
RE: You have just
)
Thanks, that's good info. Unfortunately, on a 1,1 I'm limited to two 6-pin connectors unless I want to exceed spec limits on power draw. Since those 6-pin connectors go through the motherboard, the idea scares me.
I bought a 750Ti and discovered that it won't display video on a Mac Pro, or at least a 1,1 Pro. However, it is recognized by El Capitan in the system profiler and does crunch, so I now have it paired with the GTX 950. The 750Ti looks like it will come in about 10% less output than the 950, much of which may be due to it being in the x4 gen 1 slot, which is extremely low PCI-E bandwidth.
Since a 960 is the same as a 950 unless you use nVidia Inspector to remove the memory throttle (which can't be done in OSX), it appears the best nVidia combo for a 1,1 is dual 950's. I can live with a 950 + 750Ti for now. Both are low-power cards that will easily be resellable when I am done with them.
Editing the title since it's
)
Editing the title since it's no longer an "I need help" thread.