>>48211147In fully loaded compute workloads the 290x itself can technically be bandwidth starved crunching DP. If you do the math you'll find that at 1Ghz a 290x would be trying to push 512GB/s when it has 320 available (@1350 mem clock)
With a suspected doubling of ROPs at the same clock FijiXT would still, theoretically, be bandwidth starved in DP work and only HBM2 would alleviate that.
Inb4 "theoretical max blahblahblah"
Properly coded compute work can easily reach within >95% average utilization.
You need to think about why every GCN arch has specific optimum core-to-memory clock ratios for compute work, and more importantly how that ratio effects executed code versus resources available
tl;dr you talk like a fag and your shits all retarded
