Basically I am a computer system builder & need to build a machine for a customer that needs lots of computational power. I need to find the best system for him (at same time keep the cost down) & dont know much about CUDA so any other help would be greatly appreciated.
8800 Ultra and Tesla C870 are NOT equivalent. The memory bandwidth is 103.7GB/sec for the Ultra and 76.8 GB/sec for the C870. There also seem to be various flavours of Ultras out there with different memory clocks. You will notice the difference in memory bound computation.
Does anyone have a C870 ? Where do you get 1 in the US ?
Also does anyone have a custom built system with C870 ? Because I am looking to build something & would like to know what parts I should get, motherboard, Video card, etc so I don’t run into any issues.
I tested it together with a Quadro NVS 290 and with a Geforce 8800 Ultra on the Intel S5000PSL (PCIe 8x) and the Intel S5000XVN (PCIe 16x) mainboard. Both work quite well. Those are server boards though - dual Xeons where a requirement for me. As far as I know you should be able to use any mainboard with a PCIe 16x size slot.
I have used various Nvidia Cards with the C870 both windows and Linux. as long as they are supported by the driver. This includes FX5200 and GeForce 6200 PCI cards so all my PCI express 16 slots were availble for computation. I found the GeForce 6200 PCI card with OpenGL2.0 support worked better than the FX5200 for the demo’s included in the SDK that had graphical output. This included 570 590 780 and even the AMD spider motherboard chipsets. I imagine the PCI express x1 slot cards could also be alternatives. I have had some trouble configuring cuda cards under linux with motherboard graphics adapters. I don’t know whether this is a bios problem or what. This was with an Intel chipset on a board with limited slots so I haven’t attempted other solutions.
I thought it would be best to list down my observations here.
I tried installing the C870 on a HP Compaq DX7200M Desktop system that has following config
HP Compaq DX7200 Microtower
Intel Pentium P4 2.8Ghz 1 GB Cache
1 GB Memory
Onboard Intel 945 Graphics
Linux RHEL 4.5 64-bit system.
I installed the latest CUDA 1.1 driver, Toolkit and SDK.
The non-graphics SDK samples like deviceQuery, matrixMul, etc. worked fine.
But when running a OpenGL samples fluidGL there is an error (something like “GLX extension is not supported by the display 0:0”).
We had lot of problems getting the fluidsGL and similar applications running with Intel Graphics. Finally we installed a Nvidia Quadro NVS280 PCI board and connected the video to this card. The moment I installed this card, the system started giving messages about the power starvations.
PLEASE NOTE YOU NEED A HIGHER WATTAGE SMPS (POWER SUPPLY) when the system has a C870 and an additional Nvidia card installed in the system. I replaced my 250W SMPS with a Tagan BZ 800W SMPS. These SMPSs are NOT cheap.
The Tesla C870 Installation guide does not mention anywhere about the higher wattage power supply requirement.