Well... a seed at least. And then they are expanded using AES encryption IIRC (which "shouldn't" be breakable, and even if it were breakable it'd probably be very difficult to follow). I think RDSEED takes hundreds (or nearly a thousand) cycles to complete, but we're still talking millions-of-bits of entropy per second. More than enough to shuffle a deck even if you're taking a fresh RDSEED every single card.
Every few months, it feels like "someone effed up RNG" becomes an article. But in practice, RDRAND / RDSEED are the primitives you need. And you should be getting that for free with Linux's /dev/urandom on modern platforms.
----------
I think RDSEED / RDRAND cannot be "proven secure" because of all the VMs we are running in practice though. So its something you need to be running on physical hardware to be 100% sure of security. So its still harder than it looks.
But its not "impossible" or anything. Just work to cover all the little issues that could go wrong. After all, these RDRAND/RDSEED instructions were created so that we can send our credit card numbers securely across the internet. They're solid because they _HAVE_ to be solid. And if anyone figures out a problem with these instructions, virtually everyone in the cryptographic community will be notified of it immediately.
---------
EDIT: I should probably add that using the shot-noise found in a pn-junction (be it a diode or npn transistor) is a fun student-level EE project if anyone wants to actually play with the principles here.
You are basically applying an amplifier of some kind (be it 3x inverters, or an OpAmp, or another NPN transistor) to a known quantum-source of noise. Reverse-avalanche noise from a Zener Diode is often chosen but there's many, many sources of true white-noise that you could amplify.