Can any of the NVIDIA developers with access to a 9800 GTX post what device-to-device bandwidth is reported by bandwidthTest in the SDK? We’re curious how this new card actually compares to the old 8800 GTX in memory throughput.
well you can calculate quite easily.
If you take the memory clock * 2 * buswidth / (8 * 1024^3) You will have GiB/s. The calculated value for 8800GTX is really close to the reported value. The memory clock should be around 1000 MHz, if it is around 2000 MHz, leave out the 2 in the above equation.
I have done the calculation for 9800GX2, and that one is quite slower I am afraid.
If you do this for the 8800 GTX (900 MHz DDR2), you get 86.4 GB/sec or 80.5 GiB/sec. But if you run bandwidthTest, you only get 70GB/sec in practice.
The calculation for the 9800 GTX (1100 MHz DDR2) gives you 70GB/sec on paper, but I’m curious how close the practical speed approaches that limit. If the same efficiency applies to the 9800 GTX, then that’s 56.7 GB/sec, but maybe the overhead is reduced now. Hence my question. :)
Hmm, I had in my memory that the difference was not that big, but 80 vs 70 is indeed quite a difference.
the 9800 x2 is a very nice card but if your applications / kernels are memory consuming I don’t think this is the card to go with. Because it uses the same bandwidth as the 8800. And they try to pump the data of two cards over 1 slot instead of the normal 2 or 1 on 1
The bandwith of the memory is lower than 8800GTX.
But bus-speed is (in theory) two times as big (PCI-E v2 vs v1), so getting data to the cards should not be much more trouble than with 8800GTX.
But for me (and my memory-bandwith-bound kernels) the 9800GX2 is of no use unfortunately