Actual Evaluated Throughput of TB2

[ibanez_bass]

Has anyone ran the CUDA Concurrent Bandwidth Test on their NVidia GPU (in Windows)? Strangest thing, my 2014 MBPr 15" w/ 750m gets around 850 MB/s throughput from Device to Host and vice versa with a combined of approximately 1300 MB/s. Not bad, that's around 10 Gbps combined, which is about 2/3 of what I would expect. So I ran it again, this time with the eGPU connected to my Dell XPS 13 9350 with a Thunderbolt 3 to Thunderbolt 2 adapter. This time, I got around 1300 MB/s DtoH and HtoD, and about 1800 MB/s combined, which brings it to around 14-15 Gbps combined. That's 50% more bandwidth in host to device and ~30% more bandwidth in combined. It makes a difference for sure. Can someone else run the Concurrent Bandwidth Test on their GPU and report back results?  I'm wondering if my MBP is negotiating to 10 Gbps in Windows for some reason.

[Morv]

PCIe devices connected via Thunderbolt are only running with their max. PCIe specifications compatible bandwidth. That's why with TB2 you "only" use 16 Gbps(PCIe x4 2.0) of the 20 Gbps the TB2 connection can provide. With TB3 it's 32 Gbps of 40 Gbps because of the same reason(PCIe x4 3.0). Your values are fine.

Edited August 6, 2016 by Morv

[ibanez_bass]

2 hours ago, Morv said:

PCIe devices connected via Thunderbolt are only running with their max. PCIe specifications compatible bandwidth. That's why with TB2 you "only" use 16 Gbps(PCIe x4 2.0) of the 20 Gbps the TB2 connection can provide. With TB3 it's 32 Gbps of 40 Gbps because of the same reason(PCIe x4 3.0). Your values are fine.

 

My chassis is only TB2. The values should be identical between the MBP and the XPS 15 unless something on the MBP is leaching PCIe throughput. I should be getting 16Gbps on both.