News Gigabyte introduces Thunderbolt 5 PCIe 4 card with up to 120 Gbps of bandwidth, support for 100W power delivery

[Li Ken-un]

Divide your b/s by 1024, several times, instead of 1000.
Or see my first post:

That makes sense: 4040.22MB/s = (/1024) 3.946GB/s = (x8) 31.56 Gbps

That makes sense, to me, that the speed stays below 32Gbps. And with MiB/s instead of MB/s, that works (many people are confused with them, even professionnal or company, sometimes).
But I'll give you the benefit of the doubt. Maybe it's really MB/s and so, above 32Gbps. In any case, it's well below 40Gbps. Enough room for the protocol's stuff.

What leads you to believe that connectivity speeds units should be converted using a factor of 1,024? The industry standard is solidly decimal-based here.

This is quoted straight from the USB4 specifications, Section 6.1.2.1 Raw USB Bandwidth:
(td;dr: 40 Gbps = 40,000,000,000 b/s)

The maximum USB4 bandwidth of a Link is the raw bandwidth produced by the operating link rate and the number of Lanes used. Table 6-2 shows the raw bandwidth for each type of USB4 Link.

Table 6-2: USB4 Raw Bandwidth​

USB4 Link​	Raw Bandwidth (USB4)​
Gen 2x1	10 Gbps in each direction
Gen 2x2	20 Gbps in each direction
Gen 3x1	20 Gbps in each direction
Gen 3x2	40 Gbps in each direction
Gen 4 Symmetric	80 Gbps in each direction
Gen 4 Asymmetric	120 Gbps in one direction 40 Gbps in the other direction

Note: The prefix Gbps represents 10⁹ bits-per-second and not 2³⁰ bits-per-second

CrystalDiskMark couldn’t possibly be showing raw throughput in violation of the OSI model, and Occam’s Razor would suggest that wouldn’t be a good explanation anyway, so the best explanation is that those benchmark figures represent the goodput, which get very close to the 40 gb/s bandwidth after accounting for overhead―all of which is under the aegis of PCIe.

You can also prove it to yourself (I did it with an Intel Optane P5800X) by attaching a PCIe Gen 4 × 4 SSD to CPU lanes and running it in PCIe Gen 3 ×4 mode. The SSD will be capped at 3.9 gB/s. If USB4 v1.0 were limited by 32 gb/s of PCIe bandwidth, then it could not possibly hit 4.0+ gB/s.

If the overhead were factored in with your 32 gb/s cap as a starting point for PCIe traffic, then CrystalDiskMark could not exceed 3.3 gB/s.

In fact, Intel defines Thunderbolt 4 to have a minimum PCIe bandwidth of 32 gb/s. Why would such a minimum be defined if it could not be exceeded?

I rest my case.

 

[Hartemis]

Of course. The maximal bandwidth of USB 4 40Gbps is 40Gbps, not 32Gbps. PCIe is only a subprotocol tunneled by Thunderbolt/USB4.

You took them with a grain of salt, but the screenshots from the kickstarter project are not fakes. They are possible, since they don't exceed the USB4 specifications.

I even thought they were staying below the speed of four Gen 3 lines. This isn't the case, you have demonstrated it. But definitly, these screenshots are real, and the official USB 4 40Gbps bandwidth is respected. Both statements are true. CrystalDiskMark never surpasses it.