That is very likely the result of a doubling of FP32 units like Ampere.
This should result in a significant drop in the 'Teraflop per watt' performance vs. RDNA2, much like the RTX 3080 might have 2.8 times as many Teraflops but perform 1.6 times better on average (only ~57% of total scaling).
Edit - but even without the doubling that'd be 46 Teraflops for ~ 600 W, so yeah. The thing's massive.