AMD's "Bulldozer" is no revolution

Reviewers who have got their paws on AMD information on Bulldozer and Bobcat at Hot Chips, have been giving the architecture a pretty lukewarm reception.

It is not as if it is a pile of dog poo, it is just as Jon Stokes at Ars Technica,  put it it,  more of an evolution of what has gone before rather than anything completely radical.

AMD is hoping for a lot from the chips, which are aimed at the server and mobile markets. After all they are the biggest departure AMD has made since the original Opteron. Bulldozer in particular has been claimed to be the biggest departure from existing hardware since AMD introduced the original K7 back in 1999.

While AMD claims that Bulldozer is that it’s a “third way” between traditional multicore and simultaneous multithreading (SMT) that is not exactly true.

What was presented at Hot Chips is really a new form of SMT.

With SMT, the processor keeps two or more threads loaded into the processor and the execution hardware switches between them. If a thread stalls there are instructions from a different, non-stalled thread in the machine that can be run instead.

But this form of SMT adds a bit to a processor’s die area and uses juice. It works really well on multithreaded workloads.

AMD argues that a single two-way SMT core works as well as 1.3 regular cores, because the threads don’t wait on the memory but on execution resources. The problem is that if the memory is not the cause of the bottleneck then it is likely to be the execution unit that is slowing everything down.

Bulldozer speeds this up by replicating the integer units, so that there’s one register file and one complete set of integer units per thread.

While it is a good SMT design, it is not dual core. It is more like a 1.5-core design. Good, but no revolution here Mr Guevara.

But AMD insists that a single Bulldozer core can execute two threads like a 1.8-core. We guess it might, if the wind was behind it and it was going downhill. It really depends what work the chip is trying to do.

The problem is that while Bulldozer is being marketed to such high expectations, it is possible that AMD is going to hit a problem when it fails to deliver. That is not saying that the chip is bad, far from it. But it does mean that AMD might be hyping it far too much.

Bulldozer appears to be a good conservative evolution which is a safe bet. True, it adds lots of interesting stuff, but none of it is going to be cure for cancer.

It is also not clear if Bulldozer will cause any damage at all to Intel’s Xeon line. At the moment Intel has sewn up the server market and Bulldozer might have to be slightly  better to even scratch its paintwork.