Hello all,

I have been wondering whether 8x line bottlenecks GPU crunching. I have been using GA-X58-UD3R installed with three 5870 and haven't encountered any bottleneck despite the additional third GPU, all running at x16-x8-x8 lines.

Is x8 line enough for 5870, or even 5970? What is the conversion logic behind the computer power and motherboard line distribution?

Also for further expansion, I'm considering to buy Asus X58 Supercomputer that gives 8x-8x-8x-8x if installed four gpus. It seems quite perfect.

Thank you for the replies.

The only time the bus is used is while it is loading the work, after that it is all run on the GPU itself.

People run on x4 with no problems as well.

---

I'm running 1 Box that has a 1x PCI Slot on it or at last that's what GPUz Reports with no problems Arif. It runs the Wu's just as fast as a 16 Slot does ...

---

STE\/E

The only time the bus is used is while it is loading the work, after that it is all run on the GPU itself.

People run on x4 with no problems as well.

Does such process apply to other gpu projects as well? For example DNETC@Home? I thought other than core clock, it was about the pci line bandwith that shortens the wu duration.

I thought other than core clock, it was about the pci line bandwith that shortens the wu duration.

Usually in BOINC x4 x8 x16 makes no odds, I am not aware of an application that takes the bandwidth.

Its the GPU and CPU that count. With the high end card - 5970 - to get no cpu bottleneck an i7 is needed. Will run ok on Phenom2 small hit, but not much. Lower than Phenom2 then you will get CPU bound on a 5970 slowing it as it feeds the CPU faster than the CPU processes it at times.

With PCIe standard now common V2.1, each lane on the PCIe bus has a bandwidth of 250Mb/sec each way. So a x8 slot has 2Gb/sec each way. Its bandwidth, not speed thats measured - x4, x8, x16 slots all run the same speed, but have different bandwidth. Thats where confusion sets in, assumption is made that the rating is speed, its not, its bandwidth.

Therefore for us there is little difference in practice as all the apps we run in BOINC hardly fill one lane, let alone 16 lanes, in normal useage. For sure, some differences will occur, but they are so infrequent and minor as to be ignored with safety, we rarely fill the Lanes. Your four x8 slots will be fine, no issues. In practice a x1 slot would probaly do for any BOINC application today, and x4 as a minimum card and slot bandwidth will do it fine to allow for the foreseeable future.

EDIT: Just thought .... PCIe V3 is on the verge of being released in products, still not clear when, might even be out there now, its imminent. PCIe V3 will double the speed it operates at from an (effective) PCIeV2 4T/s to 8T/s. Therefore bandwidth doubles as well, but the speed the data is shovelled down each lane doubles. So depending on when you buy, it would be worth sniffing out early PCIe V3 motherboards to look at as transfer speed will double from current PCIe V2 motherboards.

Regards
Zy

Thank you all for the replies. I can now safely order the motherboard.