Originally Posted by bmgjet
Exactly my thought,
Been trying some more settings and NB/HT overclocking brings it up slowly until 2.58nb and 3ht. Then it starts going back down.
Maybe thats one of the areas PD could do with improvement.
Having a slow L2 read must be holding a few things back.
There's definitely some sort of bottleneck with the CPUNB. My guess is that the IMC can't handle RAM speeds past the point that we're all seeing scaling in frequency. When we tighten up the timings, we're actually tightening up the IMC and the RAM. I'm thinking it's a similar situation to when a GPU is PCIe limited (like running at 1x). It doesn't matter how far you overclock the GPU, it's not going to get data fast enough so overclocking the isn't going to improve performance. There's probably some part of Piledriver that can't get data from the CPU to RAM fast enough, so overclocking the RAM doesn't increase bandwidth, since something else is holding everything back.
That's just a guess though. There definitely is a bottle neck somewhere, it's the only logical explanation for why bandwidth isn't scaling with higher frequency RAM while it scales on Intel. The fact that CPUNB overclocking doesn't help is quite worrying. The bottleneck is probably something you can't fix with more clock speed, which would explain why CPUNB and RAM frequency overclocking yields very little results after a certain threshold.
Hopefully whatever it is AMD fixes. I also think this bottleneck is very, very visible in Aida Photoworxx benchmark (but I wouldn't mind if someone else could confirm it as well. The AIDA64 page says that it is designed to test for bandwidth problems, and my 5ghz FX 8350 scores the same as FX 8150 stock.
I hope they fix it in the next iteration and that we get Steamroller in 2013 or early 2014 and PD 2.0 has some improvements here. It's definitely got to be holding Piledriver back in lots of workloads.