Funny story. The Vega64 has 8GB of RAM and 4MB of L2 cache (last-level cache) across 4096 SIMD-cores. That's 1kB of L2 cache and 2MB RAM per core.
It gets worse: although there are 4096 cores, the Vega64 isn't fully loaded until you stick 4-threads-per-core (Or more precisely, 16-wavefronts per Compute Unit (256-SIMD-cores)). That means you will actually need to run 16384 SIMD-threads before the Vega64 is fully utilized.
That's less than 512kB main-RAM per core, and less than 256 bytes of L2 cache per core. Better hope your threads are sharing a lot of memory...
--------
AMD addresses this problem with NAVI / RDNA: NAVI is fully utilized at 1-thread per core. So you only need 2816 threads on the 5700 XT to fully utilize the system.
