I understand. I mean that I want "imperceptible" latency. So the takeaway from the table is:

- For a given buffer size, the lowest latency is achieved through PCI
- You achieve the smallest ASIO buffer size with PCI
- You get lower latency with USB than Firewire

Is this partly due to RayDAT's drivers or attributable to the PCI format itself?

Would Thunderbolt performance would be similar to PCI?

I'm looking at budget interfaces. However, if those interfaces come with high latency, then they'll be useless, so I'd rather invest in a better interface.

My question is whether the format itself (PCI, Thunderbolt, USB) is a strong consideration (so I would have to upgrade my motherboard) or if I can get satisfactory performance from a USB interface.

Your table seems to show that the USB device can easily do <10ms latency, so that should be imperceptible.

It looks like you may just want to trigger vst instruments via a midi controller. Because midi going into the computer is not subject to the same latency as incoming audio you will only have to suffer the output latency.

I recently did a show with my ufx connected to a Mac mini running mainstage software and the driver reported 2.2ms roundtrip latency. So output latency alone is below 2ms at 44.1 sample rate and 32 sample buffer. This delay is very near imperceptible when playing form a midi keyboard.

I have plugged a number of other brand interfaces into the same system - some class compliant and some with their own driver and in every case the driver reports minimum 1ms more for output latency alone.

Does this mean, that my numbers are wrong ?

Ok thanks for clarification.

In this context - not having your technical insight - it was not 100% clear to me whether your statement
       "What the driver 'reports' is often intentional cheating. "
was related to all interfaces or "all/most" excluding RME.

I have no HW measuring equipment, so I can only base information on such numbers
in Cubase given by Driver/Software and your fine manual.

The intention of my question was simlpy to ensure "quality", that I do not distribute wrong information.

You don't need hardware to test RTL, but the RTL Utility from Oblique.