Hi, everyone,
We can easily get the hardware value of the GPU memory bandwidth with ECC off (e.g. 177 Gbytes/sec for Tesla M2090), but how about the memory bandwidth with ECC on? Is it also reduced by 12.5%?
Thank you in advance!
Hi, everyone,
We can easily get the hardware value of the GPU memory bandwidth with ECC off (e.g. 177 Gbytes/sec for Tesla M2090), but how about the memory bandwidth with ECC on? Is it also reduced by 12.5%?
Thank you in advance!
Why not simply measure it both ways? For example, on my K20c here I measure the following STREAM bandwidth (standard double-precision variant, i.e. 64-bit accesses):
ECC off 170.5 GB/sec +/-1%
ECC on 150.0 GB/sec +/-1%
Thanks!
I was actually asking for “the hardware value of the GPU memory bandwidth”. But anyway, based on your tests, it seems that the ECC may reduce it by 12.5%, which confirms my presumption.
My understanding is the impact of ECC on memory bandwidth depends on the access pattern, so it is not possible to state a single “hardware value”. For a given app, the impact of turning ECC on or off should be easy enough to determine.