We have 27 guests online

Login Form



Remember me
Password Reminder

AMD's FX-8150 "Zambezi" - Bulldozer in Action Print E-mail
Written by Michael Schuette   
Oct 11, 2011 at 08:13 PM

Sometimes it is hard to find a beginning for an article, and it has nothing to do with writer’s block. On the contrary, there are so many things you want to say, it is just that they seem to be inappropriate. Like rubbing salt into the wound before the wound has even been afflicted. Sometimes it is just disappointment that takes over especially if there were high hopes and anticipations – even if they were against “better knowledge”. That’s kind of where we are today.

AMD’s Zambezi /Bulldozer architecture has been one of the most anticipated and novel approaches to the existing x86 concept. Building a modular CPU, streamlined towards where it counts, that is, emphasizing integer operations by doubling the number of “cores” and sharing a single floating point unit between the latter for optimal use of the available resources including footprint appears a valuable strategy. But paper is patient and theory is gray. As with all radical detours from the beaten path, there is an inherent risk that the well-planned strategy may not work and sometimes, it is as simple as an error of the transcription, like, on a scale of 1 to 5 where 1 is given as - - and 5 is given as a ++. And then, an engineer gets involved and does the math. And assumes the "--" has to be a “double-negative”. Let’s assume something like that must have happened.

To be honest, it is not that Zambezi is really bad, it is more that, given the features and core count, one would have expected a monster of performance and that is just not what AMD came up with. Rather, by the end of the day, we are looking at performance somewhere between a PhenomII 970 and a 1075T, arguably not the slowest performers in the current CPU scene but, arguably as well, no match for Sandy Bridge or any Westmere processors from the “other” CPU manufacturer.

Let’s take a walk through the new architecture and see whether we can find out where it was that things went off. First and foremost, a few clarifications on the naming conventions:

Bulldozer is a module used in the Zambezi (desktop) and Interlagos (server) architecture. Bulldozer consists of two integer cores and a shared floating point unit. Each Bulldozer core has its own private L1D but shares the L2 cache with the second core of the same module. Interestingly, the L1D cache has been reduced to 16 kB and it is write-through, which means that the cores need to rely much heavier on the L2 cache. The L2 in turn is shared between the two cores and at 2 MB one can expect the latency to be somewhat high. This means that at least in theory, there is a limitation for high performance code. On the other hand, each CPU contains four modules and this may leverage the mentioned disadvantage.

The Bulldozer concept essentially started out with taking two individual cores with their own private L1 and L2 caches and then amalgamate them into a more integrated design.

Most programs are still relying on integer operations and as such it appears like a reasonably smart strategy to emphasize integer cores and accomplish some of the real-estate and transistor count savings by reducing the floating point units to two “hemis” that are shared between the two cores. Likewise the L2 cache is shared between the two cores of each module.

Discuss this article in our forums

Last Updated ( Nov 27, 2011 at 01:47 PM )
<Previous Article   Next Article>