Max throughput of USB Serial (CDC) over USB High Speed

[cbrake]

Hi, does anyone know of the max throughput (rough numbers are fine) you can obtain over USB Serial (CDC) on USB High Speed? We are using a STM32H7 MCU. Are we limited to classic UART baud rates like 115200, or is BAUD rate ignored for USB serial when the endpoint is a MCU?

[manuelbl]

Baud rate is ignored. So in theory, you should get close to the speed limit of high speed USB, i.e. about 30 to 40 MByte/s.

Note however that it depends on your operating system's implementation of serial communication. It might not be optimized for such high speeds as the typical serial speeds max out at about 0.1 MByte/s.

Why do you want to use CDC and not a vendor specific protocol?

[cbrake]

I've been doing some testing on a SAMD21 running Arduino SW, and getting around 2.4Mb/sec over a 12Mb full speed link over CDC, so that seems pretty good! Another team member is still developing the tinyusb based solution on a STM32 with high-speed USB.

In this application, I prefer CDC as I don't need to mess with drivers on the host side. Using COBS, packets can be framed, and nanopb is working well for encoding. (see above example).

Open to other ideas as I don't have a lot of USB experience!

[manuelbl]

What's your host hardware and operating system? A Raspberry Pi?

CDC (or a serial connection) is a valid option if it is not a consumer application as the user is usually forced to figure out which of the multiple serial connection available on the host is the one for the attached device.

For communication, both CDC and a vendor interface use bulk endpoints. So unless there is an inefficiency in the host's serial communication drivers, the performance should be similar.

I'm not familiar with COBS. But nanopb can also be used with the vendor specific interface. That's also similar.

[cbrake]

What's your host hardware and operating system? A Raspberry Pi?

Pretty much any Windows/Mac/Linux system: https://github.com/simpleiot/simpleiot/releases

CDC (or a serial connection) is a valid option if it is not a consumer application as the user is usually forced to figure out which of the multiple serial connection available on the host is the one for the attached device.

That is a good point -- one of the tradeoffs. We might eventually provide a way to identify the device using USB IDs rather than serial port, which would make it a little more reliable.

For communication, both CDC and a vendor interface use bulk endpoints. So unless there is an inefficiency in the host's serial communication drivers, the performance should be similar.

That is good to know!

I'm not familiar with COBS. But nanopb can also be used with the vendor specific interface. That's also similar.

This is a good overview of COBS: http://www.stuartcheshire.org/papers/COBSforToN.pdf

COBS is a good way to delineate packets in a serial stream.

I looked up a little about the vendor class -- how do you typically implement host drivers for this? I see there is WebUSB ...

[manuelbl]

It looks as if your host code is written in Go. Thus you'd probably want to use a package like gousb as the native API on Windows, Linux and Mac are totally different. Communication looks like in this example: https://pkg.go.dev/github.com/google/gousb#example-package-Simple, i.e. get the USB device by VID/PID, open it, claim a USB interface and start sending and receiving data.

Like many USB libraries, gousb depends on libusb. I don't know if your software provides an installer. If not, the libusb installation can be a challenge for users. You will likely run into the same problems if you try to identify the serial device by USB IDs because most of the libraries that can do that also depend in libusb.

[cbrake]

yeah, I'd really like to avoid libusb and stay pure Go, so for this type of functionality (VID/PID) I may focus on Linux for now, and implement a pure Go version.

[archef2000]

Have you ever achieved more than 2.4 Mbps over the serial port with TinyUSB?