in order to start CUDA programming we bought a Sparkle GTX285 with 2Gb of ram. We wanted to install it on a Quad core but we couldn’t because we had to remove a RAID controller on a nearby PCI-E 8x (never thought the graphic card could be that big!!!).
Anyway, now I installed it on a P4 with 1.5 Gb of ram, and I was asking myself if this configuration could slow (and if so, how much) the performances of the GTX285 when employed for GPU computing.
It’s not going to slow the actual compute performance of the GPU, but the PCIe buson a P4 might not be all that fast. That’ll slow down the process of getting the data to the card, which may or may not be an issue for your code.
How so? The later generation P4s (Prescott, Cedar Mill) shipped on the LGA775 socket and 940, 960 and P30 family of chipsets support PCI-e and the 90 and 65 micron single and dual core P4s.
A quick remark, “Pentium” or “Pentium 4” is just an Intel brand that has been used on different generation of CPU that has nothing in common (except being sold by Intel!), and you could find NetBurst architecture inside (for the firsts Pentium 4) to the actual Core2 Duo/Core 2 Quad “core” architecture, and it’s even possible that Intel futurely sold CPU with them sam “Pentium” or “Pentium 4” branding using the new Core iX architecture!!!
Anyway for PCIe performance the most important thing is the chipset in the motherboard :-)
no they’re not (at least not by anyone reputable), Pentium 4 has been dead for a very long time. avid is right, Pentium 4s were always single core Netbursts, while Pentium D were dual-core Netbursts. new Pentium is just the line between Core and Celeron.
Depending on the kind of problems you are running, 1.5GB is going to be the problem. After all you do need to get the data back to the host at some point don’t you?
That isn’t an insurmountable problem. Things like mmap() were invented to accommodate situations where working datasets are larger than physical memory. Although obviously having less physical memory than the device isn’t ideal and there are performance penalties.