Ken H (kenh.delete@this.example.com) on November 2, 2014 9:54 pm wrote:
> Brett (ggtgp.delete@this.yahoo.com) on October 31, 2014 5:34 pm wrote:
> > The bulldozer core was over optimized for server use,
>
> Er.. where on earth did you get that idea?
>
> The Bulldozer core was an attempt to adapt the Hammer design to the
> quick-release highly-synthesized design path that GPUs follow.

Bulldozer is totally different architecture from hammer, even the OOE engine, heart of the core itself, is based on different principle. (PRF vs tomasulo)

Bulldozer was AMD's first totally new core after K7.

They understood they need to move on from those old K7 based design, which were starting to lag behing intel's core2, but they got couple of things wrong.

1) They were aiming for really high clock speeds, but ended up having some bad critical paths that make it not clock so high as it should have clocked, and requiring more voltage->more power to get even those lower-than-planned clocks. Not using enough custom logic may have played a big role in this though

2) Their cache hierarchy sucks. L1I has aliasing problems, L1D's are (too) small write-through caches flooding too big, too slow L2 with lots of writes, and the L2 latency is just too long.

3) The idea of making integer cores smaller to put more of them onto the chip is just stupid in CPU as it weakens single-thread performance, which will never lose in importance on CPUs.

> The idea was "hey, we'll stall performance for one generation, but
> then we'll be able to build all sorts of customized versions."

More like "we cant afford to do all the hand work"