The new MSI 790FX-GD70 motherboard gets a rave review in the latest Custom PC (issue 71), but disappointingly the review doesn't mention if it can support 8GPUs. Four dual-slot cards will fit OK, but that doesn't mean 8 GPUs will run. Does anyone know?
Its funny, since i joined Folding after reading Issue 70, i have started to scan new issues like issue 71 for the word 'Folding' when they do a hardware review, just to see if they mention it being a good folder. I am glad i am not the only one who started doing this
Hi Richard Thanks for stopping by. It's the "in theory" bit that concerns me, I can't find anyone who has tried the 8 GPU, even on the Stanform Folding Forum
It's the 790FX chipset - it has 32 PCI-Express lanes for graphics which are split into 4x8 on that board. It'll work like the original 790FX+SB600 on AM2+ did (we tried that in the Folding feature )
Yep, second that. It was noticeable even with the 6 GPU's (3x GTX295) we had in the 790FX-GD70, the slight performance loss per GPU, compared to having a thread per GPU, on the CPU with i7. I'd definitely pay a little extra for a quad core AMD chip if you plan to run 4 cards.
Thank you for posting that JackOfAll. I was not aware that CPU could effect the performance of a GPU. I was going to build one of the top folding rigs in Issue 70 which suggests the AMD X2 7750 Black Edition. I will instead spend a few more quid and get an X4
We're only talking a couple of hundred PPD per GPU but it all adds up. And at the time I was trying to squeeze every point possible. The GPU client only needs a small percentage of CPU, most of the work is done by the CUDA kernel on the GPU. It was noticeable that Core i7 with 8 cores (OK, 4 real cores plus 4 with HT) was the better platform. Lock each of the 8 GPU clients to a CPU core and run 2 x '-smp 4' CPU clients, locking the first to cores 0,1,2,3, second to 4,5,6,7. Run the GPU clients with a slightly higher priority than the CPU clients and PPD is maximised from both the CPU and GPU clients. On the AMD platform, (I was using a 955BE), as soon as you have more than 4 GPU clients, (and assuming a x4 processor), you have to allocate more than one per CPU core. With more than one GPU client per core, running at the same priority, you're gonna lose a little PPD as the OS scheduler switches between allocating them time on the same CPU core. It's minor, don't read too much into it. I just expect that extrapolating the numbers, with 8 GPU clients on an AMD x2, the CPU component of each, 4 sharing each CPU core, you're gonna lose a little more of the potential max PPD. I would. If you can afford the extra $$$'s for a x4, rather than a x2, and you intend to run 8 GPU clients - get the quad. P.S. One last thing that comes into play - assuming you're running a CPU smp client as well as the GPU clients, you wan't the individual CPU core loads to be as even as possible. With an AMD x4, say you run a Linux CPU folding client with '-smp 4'. That launches 4 processes, and you lock them to each physical CPU core. If you end up with one of the 4 CPU cores doing other work, ie. unbalanced, cores 0,1,2 are just running a folding process, but core 3 has extra load from other processes, you effectively limit the CPU folding PPD, to the slowest )or more loaded) core. This was a problem when I switched to the AMD platform with 6 GPU's. Two of the cores had 2 GPU clients and 2 cores had 1 each. Unbalancing the load like that, was causing the CPU client to lose another 800 PPD. Two of the CPU cores (the ones with only one GPU client plus one CPU client) were running at 50% CPU, the other two - more heavily loaded with 2x GPU clients per core, at 85%. So, unbalancing the individual core load results in wasted CPU cycles, on the least loaded cores, that would have been used by the CPU client if the load was more even between cores. Does that make sense?
There is allot of information there, so from my understand that in order to get the most PPD, each GPU has to be evenly balanced with the CPU Cores. If you have a quad CPU like the AMD and have 8x GPUs, then that will equal 2x GPU per a Core. If you went for the i7 920, like yourself, that would be the maximum PPD as you have a 1:1 ratio for CPU and GPU. I hope my understanding is right? After reading Issue 71, I noticed that a new revision of the GTX295 will be coming out soon, so I will wait until that is out before I buy my rig. After looking around and comparing AMD 4x and Intel i7 920 4x, would you say it’s worth the extra 50 quid to get the Intel? For maximum PPD, would you see a real difference?
