Slashdot Mirror
AMD Licenses 64-bit Processor Design From ARM
angry tapir writes "AMD has announced it will sell ARM-based server processors in 2014, ending its exclusive commitment to the x86 architecture and adding a new dimension to its decades-old battle with Intel. AMD will license a 64-bit processor design from ARM and combine it with the Freedom Fabric interconnect technology it acquired when it bought SeaMicro earlier this year."
Maybe the new direction is going to be heterogeneous computing. We're already seeing AMD and Intel combine x86 and a GPU on one die; maybe AMD will try to combine everything and have a couple of ARM cores for low-power tasks, a couple of Bulldozer modules for more intensive tasks, all combined with their GPU.
I remember when AMD bought ATI many years ago... everybody (including us Slashdot posters) were saying what a bone-headed waste of money that was.
Now everybody's saying AMD is really fucked except for one bright spot which is its graphics division....
I guess you take the words of Intel fanboys literally. No, the Bulldozer architecture is not hyper-threading. No, it does not mean only a slight performance gain and especially not a performance loss. I recently made 3 microbenchmarks on an Opteron 6234 (Bulldozer too). I measured the negative effect of sharing some circuits in a Bulldozer core. This negative effect varies from insignificant to small (3%, 13%, 25%). I run the same two threads on the two cores of a single bulldozer unit vs two cores on separate units. Intel hyper-threading brings 30% more performance - in the best case. The bulldozer core pair brings 75% more performance - in the worst case. How can you compare them? They are not in the same league.
The funniest benchmark was the floating point. The most frequent complaint against the Bulldozer architecture is that two cores share a single floating point unit. AMD should tell one million times that yes, they share a single floating point unit, but that is a 256 bit wide unit, which can be split into two 128 bit parts. And what is the size of the usual floating point number? Not 256 bit, not 128 bit, but only 64. In reality I measured that the two cores in a single unit processes floating point instructions almost at full speed. The negative effect of circuit sharing was only 3%, barely measurable. How ironic.