Sky/Kaby Lake "U" CPU OPI 2GT/s half-bandwidth issue
The following was uncovered in finding why the Alienware Graphics Amplifier CUDA-Z Speed was operating at x2 3.0 rather than x4 3.0 link speed.
Affected ULV "U" CPU specced systems, eg: i7-6500U, would require a BIOS fix to switch the OPI (DMI) from the low-power OPI 2GT/s mode (~x2 3.0) to the high-performance OPI 4GT/s mode (~x4 3.0). The vendor would be using the Intel FITC tool to modify that bootstrap in the ME firmware, a component of the BIOS.
A OPI 2GT/s may register the link as being x4 3.0 using hwinfo64, but when performance tested will show bandwidth equivalent to just slightly more than x2 3.0 as shown below:
| OPI 2GT/s (~x2 3.0)
||OPI 4GT/s (~x4 3.0)|
I know I'm late to the party, but I'm wondering if it is indeed the CPU<->PCH link that's the culprit. According to Anandtech, the OPI interface on ULV processors can be configured to either "GT2" or "GT4" rates, the former equivalent to PCIe 3.0 x2 and the latter equivalent to PCIe 3.0 x4 or DMI 3.0. If the Alienware is set to use the slower rate, that would limit eGPU performance.
OPI 2GT/s (~x2 3.0) rather than OPI 4GT/s (~x4 3.0) is looking to be the performance handicap
Given Dell XPS 9350 runs it's OPI 2GT/s rates which limit SSD performance to < 1800MB/s it's likely the AW13 R2 does too. A AW13 R2 NVME / M.2 SSD seq read/write benchmark would tell us but as yet I haven't found one.
Based on the following, a ME firmware mod is requires to switch on OPI GT4 mode:
Further confirmed by the following link "Changed FITC setting, OPI Link Speed to GT4 - this is the performance fix for PCIe 3.0 x4 NVMe SSDs"
I'd suggest engage Dell support requesting (likely) a BIOS fix here.
This topic needs more attention. I got an XPS 13 9360 this week with a Kabylake "U" CPU. While I have not paired it up with an eGPU enclosure, I found it has this same low-power GT2 DMI mode. I swapped the 128GB PCIe drive for a Samsung 500GB 960 EVO. This connection is confirmed to run at x4 PCIe 3.0 but Read and Write speed are restricted at 1,800MB/s. I installed Samsung Magician as well as Samsung NVMe drivers to make sure I eliminated other possible causes.
I have an HP Spectre x360, model 13-w0xx, with the dual TB3 ports. I haven't been able to test it directly with an eGPU box but I believe I'm also affected by the GT2 OPI issue. Based on the feedback I've seen from the HP reps on their support forums, I'm not going to hold my breath waiting for a firmware fix to flip things to GT4. But I still want the convenience of a single-plug docking+charging station, and I'd like to have as much eGPU power as I can pull off without bottlenecking elsewhere. Given that the machine has a 2-core i7-7500U and is hamstrung by the GT2 problem, what is the fastest GPU I should bother with? I've seen benchmarks that a GTX 1080 or 1080ti is overkill because the i7-7500U will bottleneck and that a GTX 1070 is the sweet spot, but I think those tests were run on a GT4 system. How much further do I lower the GPU bar in view of the OPI issue? I don't want to run out and get a shiny new 1070 if that's also overkill.
The interesting thing about this particular 2 in 1 is that it shares the same motherboard and bios as the 7275 model which Dell supposedly claims supports pcie x4. I'm going to check the motherboard to see if there are any DIP switches that are available, other than that there must be something specific in the bios checking for the 7275 model that allows it to run pcie x4 in GT4 OPI as there isn't anything in the BIOS that allows the change.
I am now dead convinced the nomenclature in this groundbreaking post is wrong.
While I can find a single Intel BIOS upgrade notice which indeed used GT4 as the speed of the OPI link, I can't find any such document which would use GT2 as the speed of the OPI link. Nor can I find any other Intel document using GT4 in this manner.
Instead, documents like the I/O datasheet: https://www.intel.com/content/dam/www/public/us/en/documents/datasheets/6th-gen-core-pch-u-y-io-datasheet-vol-1.pdf uses 2GT/s and 4GT/s which are way more sensible: two or four gigatransfers per second. That's indeed a measurement of speed.
GT2/GT3/GT4 is the name of the Intel graphics chip and the BIOS upgrade note writer got confused and this spread via Anandtech raising awareness to this particular upgrade.
I'd recommend re-titling the issue to "RE: Sky/Kaby Lake "U" CPU OPI 2GT/s half-bandwidth issue" and use 2GT/s and 4GT/s ongoing.
Edit: ha! https://egpu.io/ultrabook-buyers-guide-external-gpu/ uses 2GT/s and 4GT/s. Did you read my post or what happened 🙂 ?
@chx thank you for the additional research and the linked Intel document. I was not aware of this issue until I encountered the Alienware 13 R2. @xcvg was the one who pointed us to the right direction in figuring out what was the culprit. We've switched to "OPI 2GT/s" and "OPI 4GT/s" as recommended.