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.
That's just blatantly wrong, the performance loss of GPUs is very well documented and gets worse as you go towards higher end models. We're talking 30%+ loss of performance here.
Sure, but it’s probably reflective of the fact that GPUs generally aren’t PCIe-bandwidth bound. Also, TB5 and Oculink2 both use PCI 4.0 x4 links.
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.
Apple Silicon has an integrated thunderbolt controller so that should have less latency than PCs that use a discrete thunderbolt controller.
2 replies →
That's just blatantly wrong, the performance loss of GPUs is very well documented and gets worse as you go towards higher end models. We're talking 30%+ loss of performance here.
Um, I have an M3 Ultra 512GB on my desk for development. Love me some Baldur’s Gate 3, everything turned up to 11…