Dirk Broer
06-03-2013, 09:25 AM
Remember the Belgian Super Computer Fastra II (http://en.wikipedia.org/wiki/Fastra_II)?
This beast combined an Intel Core i7 920 quad-core processor on an Asus P6T7 WS Supercomputer Motherboard with 6 dual-GPU GeForce GTX 295 and one single-GPU GeForce GTX 275 video cards under CentOS 5.3. It also needed one 1500 Watt PSU and three 450 Watt PSUs to feed the system and cards, delivering 12 TeraFlop in the process and enough heat to warm the department floor during winter.
Todays evolution in graphic performance is such that this custom-built (then) state-of-the-art-2009 can be reached by any well-to-do PC enthusiast, using common PC parts from any local PC hardware store.
How? you might ask, and I will show you two approaches.
Using Intel/nVidia: Take any Socket 1155 or 2011 board, as long as it has at least three PCIe x16 slots. Take any i7 Ivy Bridge (1155) or i7 Sandy Bridge-E (2011) and take three GeForce GTX Titan cards. One 1600 watt PSU will be more than enough. To surpass the FASTRA: just add another Titan if you have a PCIe slot left -but watch the power intake, you might need another PSU when adding more than one Titan, if alone for the power connectors. Socket 2011 has the advantage of having eight slots for memory and can use 6-core i7's.
Using AMD: Take any AM3+ mobo, as long as it has at least three PCIe x16 slots. Take any 6-core Phenom II or 6 or 8-core Athlon FX and take two Radeon HD 7990s. For two of those cards one 1500 Watt PSU should be more than enough. We already have a system that surpasses FASTRA and we can keep widening the gap with 8 TeraFlop at a time by adding more HD 7990s as long as we have free PCIe slots and the PSU is powerful enough.
You can see that the AMD approach is the much more economical as it uses cheaper motherboards, cheaper CPUs and two HD 7990 (like the Titan of $1,000 each) already give 16 TeraFlop.
And to make to choice for an AMD based GPU-cruncher all the more attractive: the upcoming HD 8990 promises 10TeraFlop of single-precision performance per card........
This beast combined an Intel Core i7 920 quad-core processor on an Asus P6T7 WS Supercomputer Motherboard with 6 dual-GPU GeForce GTX 295 and one single-GPU GeForce GTX 275 video cards under CentOS 5.3. It also needed one 1500 Watt PSU and three 450 Watt PSUs to feed the system and cards, delivering 12 TeraFlop in the process and enough heat to warm the department floor during winter.
Todays evolution in graphic performance is such that this custom-built (then) state-of-the-art-2009 can be reached by any well-to-do PC enthusiast, using common PC parts from any local PC hardware store.
How? you might ask, and I will show you two approaches.
Using Intel/nVidia: Take any Socket 1155 or 2011 board, as long as it has at least three PCIe x16 slots. Take any i7 Ivy Bridge (1155) or i7 Sandy Bridge-E (2011) and take three GeForce GTX Titan cards. One 1600 watt PSU will be more than enough. To surpass the FASTRA: just add another Titan if you have a PCIe slot left -but watch the power intake, you might need another PSU when adding more than one Titan, if alone for the power connectors. Socket 2011 has the advantage of having eight slots for memory and can use 6-core i7's.
Using AMD: Take any AM3+ mobo, as long as it has at least three PCIe x16 slots. Take any 6-core Phenom II or 6 or 8-core Athlon FX and take two Radeon HD 7990s. For two of those cards one 1500 Watt PSU should be more than enough. We already have a system that surpasses FASTRA and we can keep widening the gap with 8 TeraFlop at a time by adding more HD 7990s as long as we have free PCIe slots and the PSU is powerful enough.
You can see that the AMD approach is the much more economical as it uses cheaper motherboards, cheaper CPUs and two HD 7990 (like the Titan of $1,000 each) already give 16 TeraFlop.
And to make to choice for an AMD based GPU-cruncher all the more attractive: the upcoming HD 8990 promises 10TeraFlop of single-precision performance per card........