CPU I - PiFast, Cinebench, HandBrake, Blender
The chip does indeed scale to 4.3GHz under light loads, and while the 18s time isn't impressive in the context of other CPUs, it's only marginally slower than the fastest Ryzen on test, the 3950X.
But this chip is not about light loads. It's about tasking those 128 threads into something useful. Under load, Cinebench does exactly that. Running at 3.075GHz across all cores, the 25,015 score inevitably posts our best-ever results from a single CPU. It's 47 per cent faster than the previous champ, 3970X.
HandBrake is useful in showing what happens when an application is not optimised for so many threads. Performance is some way behind the 16C32T Ryzen 9 3950X. Of course, any would-be purchasers need to pre-evaluate their workloads to see if they can take full advantage of the sheer parallelism on offer here.
In another case of not having sufficient threading from the outset, performance in Blender 2.80 was a little worse than the 3970X, suggesting that only 64 threads were being employed.
Updating to the latest 2.81a whilst using the same workload shows what happens when applications are tuned correctly. Part of this benchmark is setup, where it is lightly threaded, not playing to the 3990X's multi-core strengths, and that's why you don't see the expected massive gains from the 3970X.
Needless the say, if these kinds of projects are normal in your daily workflow, there is nothing faster out there in a 1P configuration.
