AMD has unveiled some details on its forthcoming Bobcat and Bulldozer architectures. Designed for low consumption machines, Bobcat is, in contrast to Atom, an out-of-order architecture, which should mean performance is a good deal better, AMD estimating it being able to give 90% of current performance in most cases for just half a current CPU die. To recap, Bobcat will be used in the forthcoming Ontario processors which will include two Bobcat cores as well as a graphics part: the first CPUs in the Fusion range.
On the Bulldozer side, AMD has reviewed its module design. Remember, Bulldozer includes everything you currently get in a core but with two execution blocks for integers, in the place of one. In fact, each Bulldozer module will be presented to the system as dual-core, and will be sold as such. The advantage for AMD is obvious as while performance per watt and per mm² of die should increase, the second block only adds 12% to the size of a module and 5% to the size of a quad core die. Selling an 8 core processor instead of a 4 core, for an additional 5% of die is obviously a plus for AMD, and only the future will tell if it's also a plus for us given that these won't be 8 full cores.
In detail, each Bulldozer module allows you to process up to 4 instructions per cycle, against what has been 3 for processors from the K7 up to the current Phenom IIs. AMD is thus making up the ground lost to Intel, who have had this throughput since the Core 2, and are also integrating a fusion mechanism (something that Intel have also had in place since the Core 2). Each of the integer execution blocks has 16 KB of L1 cache for data and 2 ALUs. For comparison, the Phenom II has 64 KB of L1 cache for data per core, which is twice the total for a module and this reduction is surprising. The Phenom II also has 3 ALUs per core, and here there are 4 (2*2): the number of ALUs hasn't been doubled. It will be possible for the FPU to service two threads, like with an SMT system, and they will support SSE up to version 4.2 as well as AVX instructions.
Bulldozer will have a deeper pipeline than the current Phenom II, which should allow equivalent or even higher clocks than current CPUs. Branch prediction will be improved, as will prefetch algorithms. It is of course too early to judge this architecture and we'll have to wait for the benchmarks before we draw any hard and fast conclusions, something that we won't be getting before mid-2011, though AMD themselves are talking more vaguely about 2011 without giving any detail about exactly when.
There is nevertheless some bad news, feared and confirmed: while forthcoming motherboards designed for these processors, using the current AM3r2 (AM3+ ?) format, will be compatible with current CPUs, the Bobcat and Bulldozer won't function on current AM3 motherboards, which is something that will disappoint a good few users.