Trouble is, FM2+ boards are not specified for that kind of power draw
Let alone memory bandwidth (likely why GPU clock is so meh in Kaveri)
FM2++ is, yeah, I made that one up.
A specification is just that, make a board with X spec's and it will run X cpu. Doesn't mean they can't tell mobo makes to make a FM2+ that has 50amps extra power delivery and release a new CPU. Though generally it's the validation time of these things to do it right.
There isn't a particular reason that Kaveri works on FM2+, it's just a form factor they've put the CPU on, nothing to stop them releasing a bigger gpu or more cores in it and releasing an AM3+ version.
One more obvious suggestion is that more than 2 Kaveri modules and a gpu is simply too starved of bandwidth and in need of ddr4. Intel have been pushing platforms back as DDR4 delays occur. You have to get lucky, even if you're Intel, to have the right memory, process and architecture all line up perfectly. it happened significantly more often in the past as process changes were far far more predictable/easy/on time.
The other possibility is with Kaveri + HSA(HQ and hUMA) + Mantle that we won't have a game for 2 years that is cpu limited even in 2+ gpu setups.
The main reason the GPU clock seems low is likely power. If you compare 35W mobile versions of Richland, to desktop 65W to desktop 100W. In a huge number of situations a 100W richland was offering either identical or within 5% performance of the 65W model, while using 35-50W more.... The biggest reason it looks poor compared to a Ivy or Haswell in benchmarks is in the power/performance stakes which OEM's care about a lot more, as do many others. Had Richland been released with the 65W chip as the top model, gaming would have had basically identical benchmarks, many tasks would have been within 2-3% performance, a couple things like single threaded Cinebench would have looked closer to 15% worse, but who honestly gives a crap about cinebench. But by releasing it as 65W only, it would suddenly "appear" 30-40% more power efficient.
Ultimately the 100W and even the newer 200W models only made them look bad. Thing is that 65W version can overclock to identical levels the 100W chip can. Enthusiast reviews get overclocked results, you will see final performance anyway. But in stock, non enthusiast reviews, it's suddenly power competitive with Intel, and cheaper to validate, and cheaper for OEM's to cool, using cheaper PSU's, mobo's. All for a real world performance loss off a few percent.
We all know that a overclocked Intel chip can draw well over 150W as well, and they can EASILY launch a 100W, or a 150W model, but it will look bad as power efficiency tanks after a certain clock speed/voltage. They release at 70-80W for a reason, that is the optimum performance level vs power. It's absolutely sensible for AMD to target the same ballpark power level because for 90% of sales to non enthusiasts it looks a much better chip at 65-75W than 100W, enthusiasts only care about overclocked performance anyway, which will be like Intel well beyond the stock 100W tdp....
My guess is Kaveri is simply launching in a lower TDP bracket than Richland/Trinity hit for the reasons above.
