Hopefully someone with multiple gigabit cameras might also answer, these are just thoughts on the topic and other carrier boards.
I couldn’t give a definitive answer, but basically an ethernet switch shares bandwidth and so if the cameras actually need gigabit, then a switch would not work (each camera would be fighting with the other if they operate at the same time). You’d need three ports where each port has its own controller. The mini-PCIe slot of the TK1 has only a single lane, and I doubt you’ll find any mini-PCIe NICs with two or more ports and independent controllers (and the PCIe revision would have to be at least revision 2, and most of the mini-PCIe stuff is only revision 1).
The TX1 and TX2 have four lanes for PCIe on the development board. The carrier you choose (dev. kit is just one) may change what PCIe is available. Perhaps someone familiar with different carriers would comment.
PCIe revision 1 is basically a practical throughput of around 2 gigabit per lane, and so three cameras on gen. 1 (rev. 1 == gen. 1 == v1) would not work. Also, NICs with three ports and actual controllers on each port (versus being a switch) mostly don’t exist (mostly you get port counts of one, two, or four).
You might use the built in gigabit for one camera port, along with a two port NIC (so long as it has independent controllers), but you’d have to either guarantee the NIC works at rev. 2 speeds (and a rev. 2 NIC can throttle back to rev. 1 if it doesn’t like signal quality), or get one with at least two lanes. Two lanes, even at rev. 1 speeds, will always provide around four gigabit/sec…thus any two-lane or more PCIe two port or four port NIC could work if designed with independent controllers. A rev. 2 lane has a practical max throughput of about 4Gb/s, so a four port NIC would require either two lanes or a single guaranteed rev. 2 PCIe lane. A PCIe x4 four lane would be good for low latency on a four port NIC since each port would have its own lane and would not come even close to saturating the PCIe side’s bandwidth.
Not all NICs work well on Jetson’s arm64 architecture. I wouldn’t pick one up unless you see it works with native Linux drivers using the current kernel versions. Drivers available from a manufacturer typically are only for a desktop PC (those which are simply a kernel config and build typically work with any architecture). Some newer NICs may need drivers not available with current kernel versions.
PCIe NICs can consume significant power, make sure you can live with the power consumption, physical size, and weight. The dev. carrier board does not have any kind of chassis, so you’d need to guarantee rigid mounting (especially on a moving vehicle which is bouncing around). Each NIC might have more limitations on temperature, so if this is outdoors consider that.
The driver which is used with the integrated NIC is Realtek. If you get a NIC using that same driver, then you won’t have to worry about drivers. I am not recommending this NIC, but if I were to experiment, then I might start with this:
[url]Rosewill 4 Port Gigabit Card, Ethernet Ports Network Adapter - Newegg.com
[url]Rosewill | PC Components | Gaming | Kitchen Appliances
This is a lower cost NIC using gen. 2, but it doesn’t say if it is two lane or one lane, and doesn’t say if it has independent controllers for each port or not. Mostly I suspect this has independent controllers on each port just for the reason that otherwise it’d be nothing more than a single NIC with a switch…if you could research a NIC such as this and find out that each port has its own gigabit controller which won’t compete with the next port, then this NIC might be a good starting point (perhaps the lower cost is by not having independent controllers). Since this NIC is gen. 2 capable, it might be single lane…which would be good enough provided it operates at gen. 2 (note “capable” is not the same as “will always run at gen. 2 speeds”)…if it turns out that it is two lane or more, then it wouldn’t matter if it clocks back to gen. 1 (though gen. 2 would still be better performance and latency).