- 13" Dell Latitude 7380 (FHD display)
- Intel i7-7600U
- Intel HD 620
- Windows 10 Pro (1909)
Benchmarks (external display)
Installation was straightforward and pretty much plug and play. Hot-plugging or booting with the rig hooked up both work. However I don't know what to think about the low host-to-device bandwidth I am getting. My TB3 controller is supposed to have 4 PCIe lanes but I did read about downscaled performance with Kaby/Sky Lake U processors (see here and here). Is this what I am experiencing? Still my bandwidth seems very low...
Below are screenshots from HWINFO64 which seem to show that my TB3 controller has indeed 4 lanes but that only 2 are used. Any thoughts?
Aside from the benchmarking, I am rather satisfied with in-game performance. Then again, I am comparing with the iGPU...
@wzza, It's possible Dell set the laptop to OPI 2GT/s mode rather than 4GT/s. Can you identify the internal NVMe M.2 drive and run a benchmark on it?
Thanks for your reply. The thing is I know my internal NVMe M.2 port runs with only two PCIe lanes. See attached screens. I can still run a benchmark though. What tool would you recommend? Or did you mean running a benchmark of my SSD through the NVMe-to-TB3 ADT-Link enclosure?
@wzza, Sorry to see it's a x2 PCIe connection for the M.2 slot. I don't really know of a different test to confirm whether a laptop is set at OPI 2GT/s. I recognized those speeds in your CUDA-Z Performance tab as something I've seen related to 2GT/s.
If it's not OPI then it could be Thunderbolt lanes (but I think all new Dells use 4 lanes) or it could be Thunderbolt 2 link speed instead of Thunderbolt 3 link speed (check the Thunderbolt 3 cable - is it a Thunderbolt 3 cable or a USB-C cable?).
The speeds I'm getting seem even lower than what was reported for 4 lanes with OPI 2GT/s mode.
I know for a fact that the laptop has 4 TB PCIe lanes, which is confirmed by the reading from hwinfo (that also shows that only two are used). Am I right in believing that if the OPI 2GT/s mode were set, all 4 PCIe lanes would still be used? I'm using the TB3 cable that came with the ADT-Link device and I don't have a spare to try...
On thing I will do later is have the eGPU plugged to my internal NVMe M.2 port and my internal SSD in the TB3 enclosure, and benchmark both.
I'm using the TB3 cable that came with the ADT-Link device and I don't have a spare to try...
32Gbps-TB3: to achieve ~2254MiB/s H2D (Nvidia) or ~2550MB/s (AMD) requires the correct TB3 enclosure firmware, a 40Gbps TB3 cable , NO additional active TB3 enclosure ports and a x4 3.0 host PCIe port (not x4 2.0). U-CPU systems must be in OPI 4GT/s. Macbooks with Nvidia eGPUs must not boot using apple_set_os.efi. FYI: Intel throttles 32Gbps-TB3 to 22Gbps which benchmarks as 20.40Gbps (AMD) or 18.91Gbps (Nvidia) in the important H2D direction
Apart from the OPI mode which I can't check I'm pretty sure I meet all the requirements to get max bandwidth. I'll let you guys know about my future investigations (see my post above).
Alright so here's what I did:
1) Took out my internal NVMe SSD and hooked it up through the NVMe-to-TB3 ADT-Link enclosure
2) Hooked the eGPU (R43SG) directly to the M.2 NVMe port
So I'm getting better performance using the M.2 port. Does that make sense given the fact that, on paper, it's 2 lanes (M.2) vs 4 lanes (TB3) both running with the same OPI speed?
EDIT: Found this build with a Latitude 7480 (hardware should be almost identical to the 7380) and 6600U. The reported bandwidth looks like 4 lanes in OPI 4GT/s...