Etherium is Dagger-Hashimoto, which tries to be RAM-bandwidth bound. GDDR5 and HBM2, which can achieve RAM-bandwidths of 500+ GB/s (compared to a typical CPU+ Dual-channel DDR4 with only 50GB/s), so Etherium's best machine are the cheap GPUs with GDDR5. HBM2 (Vega64 and Titan V) are faster, but I don't remember if they win in price/performance. HBM2 might win in ongoing costs due to being more power-efficient.

Monero is Cryptonight, which tries to be memory-latency bound (Ideally: executing out of L3 cache of CPUs). Unfortunately, it seems like the miners have figured out how to parallelize it, as HBM2-based cards manage to perform incredibly well in Cryptonight.

Basically, the only cryptocoin that MIGHT be worthwhile on an APU connected to DDR4 is Monero / Cryptonight. Even then, because Cryptonight is primarily memory-latency bound, I doubt that the GPU-portion would be any faster than the CPU. The majority of the algorithm is waiting for L3 cache to respond.

The L2 Cache in GCN (and probably Vega) is 512kb and shared between 4 compute units. So its too small to execute Cryptonight. As such, why execute from the on-board GPU when the CPU already maxes out the L3 cache?

Cryptonight requires 2MB of low-latency access per instance. Threadripper with 16 cores and 32MB of L3 is basically the most ideal CPU setup. But Vega64 managed to outperform Threadripper by significant margins, probably because HBM2 RAM is incredible (and there seems to be some kind of unexpected parallelization trick in the Cryptonight algorithm)