AMD Magny Cours 16-Core, Now Folding for Team 93
Leonardo
Wake up and smell the glaciersEagle River, Alaska Icrontian
16 Cores of Magny Cours fury is now Folding for Team 93. I built this machine last weekend and have been validating the hardware and BIOS this week in Windows 7 Pro. Within a couple weeks, I'll have it running in Linux and Team 93 will see a boost in production.
0
Comments
Mothterboard: Asus KGPE-D16, Socket G34 (northbridge fan is my own add-on)
CPU: AMD Magny Cours Opteron 6136, 8-core, 2.4GHz, 2X
CPU heatsink: Thermaltake, heatpipe, aftermarket fan
PSU: Corsair CMPSU-AX750
Memory: 16GB (8 X 2GB) G.Skill Ripjaws DDR3 1600
Case: 'nekkid' and happy!
GPU: integrated, on-board, A-Speed Max-Garbage Thriller (....or something like that)
A couple more images for you:
My thanks to _K_ and Tushon for advice in BIOS settings. This build was my first peek ever into a server BIOS. (lots of strange stuff there )
Are you ever going to case it, or just leave it out?
As much as he loves nice cases I bet he lets this bad boy go commando!
Leo, Congrats on the awesome build! If you are trying to lead by example again, I think you should have new followers now. I noticed in the first pic that the PSU appears to be levitating. Or is it just hovering from the brute power around it?
Mr. Goat, I will be taking my Rocketfish/Lian Li case out of retirement for this little beast. I had tried to sell that big aluminum box before, but couldn't. Well, I'm glad now that I still have it.
Honestly, that sounds about right for non-bigadv ppd. My 2.6GHz 6 core Magny Cours server running Debian 6 puts out around 10k ppd. If you scale that up to 16 cores, that works out to around 28k ppd. Of course, I do have things (like the minecraft server) running on that box as well, but still.
Where you're really going to get the ppd is chewing through bigadv workloads.
Yeah, it's about right for Windows, but there's a lot of potential that Windows can't tap. I'm not speaking as an OS expert, just from what I've seen in my forays in the Folding world.
I assume your Phenoms were folding standard (A3, non-bigadv) SMP work units.
Standard SMP work units don't receive as much of an aggressive bonus for early return as the bigadv units. The bonus awards achieved by the faster work completion and return times from Linux processing are much more pronounced with -bigadv work units than with standard SMP units.
I don't plan on running Linux for any of my 8-thread Folding machines - I7 Sandy Bridge and Lynnfield CPUs.
That makes lots of sense, Linux for the machine that is 16-core total as Linux does not care how many cores it has (it was designed for servers from day one), and Windows for the lower core amount machines (Windows now is finally (IMO) catching up with Linux as to how many cores it can use on non-server versions).
Great build and I'm excited to see how your PPD increases with the switch to ubuntu. Will send you an email about samba config for sharing
Yessiree, this beast is now Folding in native Linux, Ubuntu 10.10. Huge learning curve for me.
Outstanding performance increase so far. I haven't folded any bigadv work units yet, as I'm starting and stopping, and reconfiguring - learning Ubuntu (at least I think I'm learning). But anyway, take a look at this:
P6098 (Core A3, standard, non-bigadv, SMP2)
Ubuntu - TPF avg. 5:00, est. PPD 47,700
Windows 7 Pro - TPF avg. 6:30, est. PPD 32,500
I'm really looking forward to the Samba guidance. I've got Samba already installed, but can't for the life of me figure out how to get my my Win7 boxes to see the Ubuntu machine.
Think of linking to a limited NT Server emulator-- that is what SAMBA does. Seeing the whole Ubuntu machine is improbable, you will see areas shared via Samba, just as a more modern peer-to-peer network shares limited specific file areas and/or printers on each machine if it is set up right.
I won't hog this thread anymore, Tushon has it in hand it looks like.
Windows non-GPU client chews up CPU and RAM/CPU interfacing resources and RAM, and due to CPU resources being reserved in Windows for client with priority that can be set too high and leave CPU idle but unable to release resources at times even folding itself can be inefficient in Windows.
Linux handles process priority better, and processes that are inactive can be state-suspended while active ones get their freed resources-- and Linux is more active in doing so. This is partially an artifact of Linux being designed for servers, which have to balance more tasks efficiently than do plain old desktops. Win 7 is not even a server o/s in the full sense, so it has a different resource recovery and allocation strategy (sloppier one) than Linux does.
But(BIG ONE), what Leonardo has on his new 16-core box are also more of workstation-and-server-also designed CPUs from the bottom up(marketing viewpoint), so they are a great match for a server o/s as far as efficiency being more fully uniform in the system Leo is configuring.
From my point of view, system is software (O/S is software too) plus hardware plus people involved, and tuning Folding extremely is a great learning exercise in I/S analysis.
One more thing-- FERMI Labs is NOT folding in Windows(suspect Unix/Solaris/BSD/possibly Linux boxes in a mixture there), and they fold so fast there that Vijay had to ask and beg them not to fold too much and too fast and discourage others too much. There is no Team FERMI now(not sure there ever was one), but FERMI folds blazing fast with their boxes' spare time cycles. Vijay is leading dev into new clients away from what FERMI would find best tuned for them ("NOT FOR FERMI" is in one older Forum thread (about a beta client)'s title).
John.
P6904, TPF 53:03, PPD est. 111,745
P6901 - TPF 16:45, 71,700
P6903 - TPF 37:35, 115,000
He is also folding on Linux (better for bigadv, but worse for GPU3 to my knowledge as you have to add WINE into the mix)
1. My dedicated Folding rigs are all Linux now. You can set up GPUs to fold in Linux, but it's a pain.
2. Multi-core CPU Folding in Linux has become very productive. With respect to the compromises one must make concerning hardware purchases, power consumption, and plain old whirring fan noises, GPU Folding doesn't make sense for my Linux platforms strategy.
For Windows, GPU Folding, IMO, still makes sense. Unless you are running Sandy Bridge E or I7 9XX class CPU, GPUs can produce a large proportion of a system's Folding output. You do though, have to take heat output and power consumption under consideration. I used to fold with 9800GX2s and GTX 295s, but the power bill got a bit high and I scaled down to GTS 450s.
While we are on this topic, I've got two more GTS 450s for sale. If you want to fold with them, expect about 100 Watts each at full load; 7,200 PPD advmethods; 9,000 PPD non-advmethods.