If so, what is the performance like? Is it even supported. Is Einstein strictly single precision flops?
Hi!
We have a few GTX 680s, see the host statistics toplist here:
http://einstein.phys.uwm.edu/top_hosts.php, so GTX 690 should work as well, potentially treated as 2 separate GPUs by BOINC. [Maybe those dual GTX 680s in the toplist are actually single GTX 690s??]
At the moment, all of the GPU apps (CUDA and OpenCL) entirely work with single precision in the GPU. So double precision capabilities of the GPU are not required, and the app will not benefit from DP capabilities if they are present.
The GTX 690 is essentially two GTX 680s in one card with slightly lower clock frequencies. The GTX 680s work with this project and based on that, the GTX 690s should work also. The BRP4 application benefits from extra PCI-E bandwidth. With the GTX 690, two GPUs would share the same pipe. Based on that, the more optimal solution might be to go with two or three GTX 680s where each card has its own pipe to the CPU. The price of two 680s is also around the same as one 690.
For very good performance with this project and a high production system, I would suggest matching up the 680s with an X79 board supporting PCI-E 3.0 x16/x8/x16 slots. The Asus Sabertooth, Rampage IV Formula, and Rampage IV Extreme are good choices for this and can utilize all 40 lanes provided by the CPU. The Formula and Extreme boards support x16/x16, x16/x8/x16 or x16/x8/x8/x8 slot configuration natively. There is a simple registry modification that needs to be done to enable PCI-E 3.0 support in Windows. Either the 3820 or 3930K would a good choice for the processor depending on how many cores you want to have. I would also suggest overclocking the processor between 4.0 and 4.5 GHz using a decent cooling solution to maximize performance for the GPU tasks.
In regards to the 690 and PCIe issue. I would suspect the 690 might actually be bottlenecked even at 3.0. I say this because a 680 when running 3 tasks nearly maxes the bandwidth, and with 4 running I would suspect it would be maxed. Since a 690 could run 6 tasks at a time, I would be inclined to believe performance would be lost.
Has anyone crunched a GTX 690 on Eintsein?
)
Hi!
We have a few GTX 680s, see the host statistics toplist here:
http://einstein.phys.uwm.edu/top_hosts.php, so GTX 690 should work as well, potentially treated as 2 separate GPUs by BOINC. [Maybe those dual GTX 680s in the toplist are actually single GTX 690s??]
At the moment, all of the GPU apps (CUDA and OpenCL) entirely work with single precision in the GPU. So double precision capabilities of the GPU are not required, and the app will not benefit from DP capabilities if they are present.
Soo...anyone with a GTX 690 here?
Cheers
HB
The GTX 690 is essentially
)
The GTX 690 is essentially two GTX 680s in one card with slightly lower clock frequencies. The GTX 680s work with this project and based on that, the GTX 690s should work also. The BRP4 application benefits from extra PCI-E bandwidth. With the GTX 690, two GPUs would share the same pipe. Based on that, the more optimal solution might be to go with two or three GTX 680s where each card has its own pipe to the CPU. The price of two 680s is also around the same as one 690.
For very good performance with this project and a high production system, I would suggest matching up the 680s with an X79 board supporting PCI-E 3.0 x16/x8/x16 slots. The Asus Sabertooth, Rampage IV Formula, and Rampage IV Extreme are good choices for this and can utilize all 40 lanes provided by the CPU. The Formula and Extreme boards support x16/x16, x16/x8/x16 or x16/x8/x8/x8 slot configuration natively. There is a simple registry modification that needs to be done to enable PCI-E 3.0 support in Windows. Either the 3820 or 3930K would a good choice for the processor depending on how many cores you want to have. I would also suggest overclocking the processor between 4.0 and 4.5 GHz using a decent cooling solution to maximize performance for the GPU tasks.
Very nice post. In regards
)
Very nice post.
In regards to the 690 and PCIe issue. I would suspect the 690 might actually be bottlenecked even at 3.0. I say this because a 680 when running 3 tasks nearly maxes the bandwidth, and with 4 running I would suspect it would be maxed. Since a 690 could run 6 tasks at a time, I would be inclined to believe performance would be lost.