I know this has been discussed before, but I couldn’t find it clearly stated, so I thought I’d ask: Can anybody confirm that the K20 is limited to PCIe v2.0 in the VBIOS, hence no driver can and will be able to relieve this limitation (i.e. un-cripple it)?
I know that the K10 is also PCIe v2.0; how about the Quadro-s?
Quadro K6000 is PCI-E 3.0, so is GTX Titan. There is an upcoming Tesla K40 that is also PCI-E 3.0, but as of yet unreleased. I do not believe there is a way to ‘uncripple’ the K20, it was made to support PCI-E 2.0 speeds.
K10 is PCIe gen3, K20 is PCIe gen2.
The settings are in the VBIOS.
Thanks for the feedback!
It sounds like the K20 ended up as an unfortunate victim of hardware compatibility issues (or maybe vendor war?). I wish this limitation was not baked in into the boards and the users/compute centers had the choice of simply not bothering or taking the extra mile and ensuring that PCIe v3.0 works with the hardware they purchase.
While it is annoying that the K20 is PCI-e 2.0, keep in mind it has two copy engines which does make up for it;
[url]https://developer.nvidia.com/content/how-overlap-data-transfers-cuda-cc[/url]
Yes, it does if your algorithm allows it to be used. However, in general, there is nothing that can make up for the reduced latency and increased bandwidth of PCIe 3.0. Moreover, in any latency-sensitive application with heterogeneous paralellization this does make quite a bit of difference; the faster the GPU gets the more (proportionality) the transfers take.
Anyway, we are getting off-topic.