Oculink is generally faster than TB5 despite them both using PCIe 4.0, because Oculink provides direct PCIe access whereas Thunderbolt has to route all PCIe traffic through its controller. The benchmarks show that the overhead introduced by the TB5 controller slows down GPU performance.
It's not just the controllers; the Thunderbolt protocol itself imposes different speed limits. The bit rates used by Thunderbolt aren't the same as PCIe, and PCIe traffic gets encapsulated in Thunderbolt packets.
Maybe; I'm unable to find any benchmarks that specifically compare PCs with TB to Macs to test this. But there is certainly still overhead with TB no matter what, and therefore it'll never be as fast as Oculink.
Sure, but how big of a difference is there? Even inside a desktop PC, you typically have PCIe ports directly off the CPU and ones off the chipset, and the latency for the latter is double. But the difference is immaterial in practice.
I think latency is the wrong focal point (more important for gaming, plus Macs don't support eGPUs anymore). There aren't a lot of general workloads that require high sustained throughput, but the ones that do can benefit from TB5 scaling.
For instance, if you cluster Mac Studios over TB5 with RDMA, the performance can be pretty stellar. It may not be more cost effective than renting compute for the same tasks, but if you've got (up to) four M3 Ultras with a ton of RAM, you'll be hard pressed to find something similar.
That's still not more ideal than having native alternatives like OCuLink or something that can be networked like QSFP, but it's a fair way to highlight the current design's strengths.
