LONG AWAITED, AMD's desktop Bulldozer chip arrived roughly a quarter late compared to original expectations as July became October.
The FX-8150 processor is the top end part in AMD's initial 'Zambezi' Bulldozer line, yet it doesn't face Intel's top parts but the midrange Core i5 2500K-class Sandy Bridge. Whether this is the unfortunate result of a lack of performance in this chip, or just an initial dip in what could be a much better performing architecture in its later iterations, we will have to see over the next year or so.
During that period, new operating systems with improved thread schedulers will come along and there will be a chance to see how the next generation 'Piledriver' update of the Bulldozer fixes any initial issues.
Bulldozer's architectural premise that there should be effectively more integer core capability in the resource pool - every two integer cores share one floating-point unit - does make logical sense if you look at typical thread contents in either Windows or Linux. There are many threads in some workloads that are purely integer, and saving valuable die space by not having a floating-point unit for each core handling such threads allows more overall cores on a die.
However, issues where those two cores are supposed to simultaneously share that floating-point unit and then quickly switch to a mode where only one of them has full control of that same floating-point unit can lead to various problems, including performance. Remember that Bulldozer has deep long pipelines, as well as long latencies for cache and memory, in order to allow higher clock frequency, which in turn increases the register save and restore penalty when switching modes.
This is akin in many ways to, guess what, the Pentium 4 Netburst microarchitecture. If you remember this, you will also remember that Netburst hit a wall quite quickly, and was replaced by a slower clocked but much lower latency and higher real performance Core microarchitecture.
My personal preference is that the floating-point unit sharing mode be done away with, and simply have the second core as purely integer. This would also simplify the associated OS thread schedulers, as they would simply use an even-odd approach, with every 'even' core having floating-point and every 'odd' core for integer tasks only.
Now, let's look at the machine. We surrounded the FX-8150 with the best there is for it: the Asus Crosshair V Formula mainboard and AMD's Asetek-made liquid cooling kit, as well as a G.Skill RipJaws X 16GB memory kit composed of four DDR-1600 4GB DIMM modules. The Asus top of the line AMD mainboard with the 990 series chipset in fact costs more than this CPU, a reminder that AMD has to find a way to charge more for its desktop CPUs, which will only be possible if their performance goes up, obviously. The G.Skill high capacity memory bundle occupies all the memory slots for maximum board capacity, and even colour matches the board.
Sign up for INQbot – a weekly roundup of the best from the INQ