Electronic – Uniquely Identifying FT232H Chips

Yes, you are indeed filling up the FT245R's 256-byte transmit buffer.

Remember, USB is fundamentally a half-duplex communications protocol. If you send hundreds of bytes to a device in a burst, there's no opportunity to receive bytes back from the device until the burst is over.

Your CPLD is having no trouble accepting data from the 128-byte receive buffer, keeping it from filling up, at least until the transmit buffer fills up. But with a total transmit capacity of 256 + 2 bytes of buffering, any burst over 258 bytes is likely to fail.

There is some "randomness" in the results with hundreds of bytes because of the fact that Windows polls the device status every 1-2 ms. If it happens to poll the device in the middle of a large burst, before the FIFO has filled up, it will drain the available data from transmit FIFO, making room for another 258 bytes before it fails. But at a peak transfer rate of 1 MBps, any transfer longer than about 500 bytes is pretty much guaranteed to fail.

Bottom line is, if you need to transfer large amounts of data in bursts, you need to have sufficient memory for buffering those bursts outside the FT245R.

If you can use your CPLD to set up a 2-KB data FIFO (perhaps using an external chip), your stress test will work perfectly with any amount of data.

As Matt says, the first thing that jumps out is more capacitance needed.

Breadboards are terrible for stray inductance/capacitance so you really need to try and keep wires as short as possible, keep sensitive signals away from higher current/voltage/fast changing signals, and add plenty of bypass and bulk capacitance around the circuit.

I would start with a couple of >100uF electrolytics on each power rail and near the FET, with a few 100nF or higher ceramic caps on the power pins of the ICs.

Another thing to look at is using a separate rail for the relay, and routing the ground return separately from the uC's ground. Of course they need to be tied together, but you can ensure the high current return does not flow through the uC board's ground this way.