Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Thunderbolt 5 data transfer between two computers
- Thread starter Xenobius
- Start date
- Sort by reaction score
About 65Gbps between M4 Max MBP and M4 Pro mini
Better than I was expecting.
Better than PCIe 4.0 x4 (63 Gbps).
If Intel PCs are using PCIe 4.0 x4 Thunderbolt 5 host controllers (JHL9580, JHL9480), then they can't achieve 65 Gbps. You need PCIe 4.0 x8 or PCIe 3.0 x16 to achieve more than 63 Gbps. Or they need to integrate the Thunderbolt controller into the CPU like they did with IceLake and TigerLake or like Apple is doing with Apple Silicon.
Regarding SSDs, those will be limited to 63 Gbps regardless of Mac or PC because PCIe 4.0 x4 is the downstream PCIe connection.
Unless someone makes a Thunderbolt 5 (or USB4 v2) peripheral controller that is not limited to PCIe 4.0 x4 downstream.
For SSDs, Mac Thunderbolt 5 could be slightly faster than the discrete Intel Thunderbolt 5 controllers because the 63 Gbps limit exists only at the device end instead of both ends.
Unless someone makes a Thunderbolt 5 (or USB4 v2) peripheral controller that is not limited to PCIe 4.0 x4 downstream.
For SSDs, Mac Thunderbolt 5 could be slightly faster than the discrete Intel Thunderbolt 5 controllers because the 63 Gbps limit exists only at the device end instead of both ends.
Many thanks.
In a test like this, isn't the transfer rate limited by the speed of internal SSD?
Could you mount and use a RAM Disk on both computers instead of the SSD and repeat the test?
Could you do a test with bi-directional data transfer at the same time?
Oh I was just posting Alex Ziskind‘s short test video, that wasn’t done by me. You can perhaps ask him directly on alternate test perimeters.
The iperf3 in question disregards storage speed and only looks at packets, so the bottle neck is likely to be in the NIC than anything else like CPU, RAM or the rest of the system.
And ~64 Gbps looks like it is already the best case scenario in a 80Gbps bi-directional data lane (doubled from TB4’s 32Gbps).
63 Gbps is 7.875 GBps.
File sharing bandwidth will be less than the iperf speed due to file sharing overhead.
Transfer rate will be limited by the slowest device between each computer.
A RAM disk could work to maximize file sharing benchmark results (if the internal drive is not fast enough), or this:
A Mac mini has three Thunderbolt 5 ports. Connect a Thunderbolt 5 software RAID to each Mac mini to get above 80 Gbps (or at least above 63 Gbps). Use two different Thunderbolt 5 ports for the software RAID. Then see how fast that data can be put through the third Thunderbolt 5 port to the other Mac.
I wonder what people would use Thunderbolt for if it had PCIe 6.0 x4 instead of just 4.0? Some companies like Sonnet Tech have external PCIe enclosures. Wouldn't mind seeing Macs able to use higher end PCIe cards & newer M.2 cards at their full (or at least fuller) potential.
For Thunderbolt 5 120Gbps (15 GBps), PCIe gen 5 x4, gen 4 x8, or gen 3 x16 would be sufficient.
120 Gbps is the asymmetric mode of Thunderbolt where one direction can do 120 Gbps while the other direction is limited to 40 Gbps.
Can Thunderbolt 5 negotiate 120 Gbps transmit when doing writes and 120 Gbps receive when doing reads? That would be complicated. Maybe 120Gbps is only used when sufficient DisplayPort bandwidth (transmit only) is requested.
Should read the USB4 v2 spec to find out what kind of control the OS can have on the link width and if the host controller has any link width automation.
Not in the way you think.
A Thunderbolt peripheral needs one and only one Thunderbolt host, since it acts as a PCIe extension from the host computer. So whichever Mac that is first that the hub sees will become the one host.
The other 3 Macs afterwards when plugged into these ports, at first they won't do anything since the Macs are not peripherals to the first Mac. However, if you at this point engage Thunderbolt network bridging modes on each of these Macs, they now then can talk to each other over a p2p TCP/IP network. I know that one to one connection must work, but I am unsure if the TB hub can act as the network switch in this case, or is the host Mac going to do it by proxy.
As I understand it, yes, 120 Gbps mode is exclusively for situations when DisplayPort is signing up to statically consume a lot of TX bandwidth.
I haven't read up on the details but I would be surprised if it's capable of switching of which side gets 120 Gbps TX on the fly without substantial performance loss. Every direction change requires shutting down one of the three transmit lanes of the current 120 Gbps side, then bringing up the corresponding transmitter in the other side. That's going to require some link training (sending test patterns and using feedback from the other end to tune up the various signal equalizers along the chain) and a re-do of lane bonding. Unless Thunderbolt link training is unreasonably fast, doing this frequently is not compatible with high performance.
(In practice I'd bet that when higher level software decides it's time to reconfigure for 120 asymmetric mode, they simply shut down the whole link and re-run link training and bonding from scratch.)