I think the whole SMT, CMT, and VMT situation has been confused.
Bulldozer/Piledriver/Steamroller have parts that are VMT.
Bulldozer/Piledriver; Fetch/Decode is VMT and so is the front end for the FPU.
Steamroller; Fetch is VMT and so is the front end for the FPU.
Excavator; Everything is SMT, there is a fetch for each core, there is a decode for each core. While the front-end for the FPU can receive instructions from both cores at the same time.
Excavator gives up the VMT structure that is in Steamroller, Piledriver, and Bulldozer for SMT.
Originally Posted by Usario
Four integer clusters in a module? What are they going to do, double up the scheduler again? And then leave it with a 4:1 integer:FP core ratio, or double up on the FPU as well (and would that mean having two separate FPUs or a 512-bit FPU that builds upon the current Flex FP to do two 256-bit ops, four 128-bit ops, etc, concurrently)? I'm no expert but ditching CMT altogether sounds better to me.
The current implementation is;
One core has one execution/memory cluster.
The future of CMT is CCMT which is;
One core with two to eight execution/memory clusters.
The FPU in Bulldozer/Piledriver/Steamroller/Excavator is not part of the core but part of the module. AMD will keep the FPU consistent with the capabilities of the x amount of clusters.
So, if you have a CCMT module with 8 cores and each core has 8 clusters; then you'll have 64 x-bit FPUs.Edited by Seronx - 5/4/14 at 2:29pm