Well before Bulldozer had taped out, AMD briefed us on the high level processor plans (our software was included in BAPCo Sysmark and we'd been a very early adopter of AMD64, so we had lots of these meetings). I told them that it was a total mismatch for our workloads (audio and video production tools) and would prove to be a bad fit. They protested at first, as our inputs and outputs were fixed point formats and the doubled-up integer/logic units would allow us to thread more. That's when I spelled it out:

1. Our business logic and UI each basically operated in single threads.

2. Our intermediate audio and video processing buffers were floating point or moving in that direction in the future.

3. Wherever possible in filtering and effects, we avoided branching, relying on zero-folding in vector instructions.

I went on to explain where they'd be in trouble for game engines, document processing, and other key use-cases.

Their hardware architects slowly sank in demeanor over the course of that all-day meeting. I later heard from contacts at AMD that it was an "oh shit" moment for them. That said, having been at Sony at the time, we'd gone down the long dark road of the Cell...

Sometimes people try bold and interesting things that don't work. Bulldozer was definitely one of those, and it should have been killed (or at least diverted), like Larrabee. Nonetheless, AMD survived, and their current architectures owe some of their thinking to the hard-learned lessons of Bulldozer.