My personal bet is that they're running a custom APU that is a Zen 2 chiplet with a full Navi die on the same package. They're clocking the hell out of it to match final retail TF. This allows them to get a close to final configuration with more realistic on-package latency between the CPU and GPU.Maybe the devkit isn't anywhere near final and is running discrete CPU and GPU - i.e. not final silicon - and doing so with the closest equivalents, the GPU being 5700XT clocked at 2GHz+ to get to 10TF+. That would require some beefy cooling.
Final hardware then goes a bit wider and slower - anything from 44CUs @ 1.8GHz to 52CUs @1.6GHz to get the in same kinda sorta envelope.
This way, it will appear in all of these performance loggers are a complete part and give the impression it could be ES silicon of the final hardware, when the true APU is only just entering production.