WD_Black D50 Game Dock Thunderbolt 3 NVMe M.2 Docking Station
I have seen Sonnet's video previously and it is an excellent explanation of the differences between Thunderbolt 3 and Thunderbolt 4.
I am interested to see how the Lenovo Thunderbolt 4 dock performs with an eGPU in a daisy chain. It uses the JHL8440 Controller and has a single Thunderbolt downstream port, which according to Sonnet's video can support the passthrough of 4 lanes of PCIe.
Ive recently tested daisy chain TB4 laptop > TB4 dock > TB3 eGPU with the kensington SD5700T (SKU: K35175EU) and it works but with similar performance loss as with TB3 dock.
But I don't know what controler this TB4 dock uses
I expect I will have the same results but I am interested in seeing how it compares to my old Dell WD19TB Dock's passthrough performance loss.
@boltoway, is it the JHL7440 or the firmware on the JHL7440? I wonder if there's a setting in the firmware that can be used for the same effect on an Alpine Ridge? We don't have a utility to dump all the registers yet. Linux can dump the contents of the registers but doesn't label each field (there's more than 3000 of them listed in the Alpine-Ridge_DP_1.03.pdf document).
Maybe the controller isn't everything! My Chinese Domestic Market eGPU with a JHL6340 just outperformed by Akitio Node Titan, even though it reports a lower H2D speed:
Akitio Node Titan (JHL7440):
CDM eGPU (JHL6340):
The margin isn't very big, but it is not the result I was expecting.
results are in the marging of jitter / error, cooling and background tasks can be involve in those variance.
On timespy graphics score are almost the same, it's CPU score that make the difference.
That said, not having the JHL7440 above it's surprising.
@boltoway, your D2H number has the greatest difference but I think H2D is more important (being able to send data to the GPU faster).
Lower H2D might mean lower latency - which might be better for some GPU tasks. There is a tradeoff between bandwidth and latency with Thunderbolt and USB4 (the USB4 spec mentions this). Maybe eGPU vendors started with low H2D but changed their firmware for high H2D because people complained (H2D is easier to show in a benchmark than latency). Maybe this tradeoff is one reason why eGPU enclosures have separate certification than PCIe/NVMe enclosures.
I think we need tests with different firmwares to show how H2D relates to benchmark scores for Thunderbolt.
A higher H2D might be better for min FPS score because this is when the GPU needs more data.
A lower H2D (better latency) might be better for max FPS score because this is when the GPU needs more draw commands from the CPU.
Yeah the scores are definitely too close to be conclusive. In saying that I am going to run a series of benchmarks and compare the average. The CDM eGPU appears to be fairly consistent with results which is good, although it also appears the Silverstone ST55F-PB I have running it isn't powerful enough to run my Gigabyte 3060 Ti Gaming OC Pro. Each now and then during Timespy the GPU and PSU will shut down, leaving only the BUS powered Thunderbolt Chip running. I have reduced the power limit to 80% which seems to have stabilised it for now, but a more powerful PSU is in order.
That's a good point. It would be interesting to be able to flash different firmware onto different controllers and having a normalised test with the same GPU and host to compare results.