AMD leak confirms that Excavator APU will be 28nm, and that some production … – ExtremeTech
Over the past few months, hints and tidbits about AMD’s next-generation Carrizo APU have begun to leak into the market. One of the most fundamental questions has been which process AMD would target for the part — 28nm or 20nm? Timing suggested that AMD might go for 20nm, but new information suggests that’s not the case. A reader tipped us off to a forum post over at SemiAccurate where someone noticed LinkedIn data suggesting that Carrizo, AMD’s first APU with an Excavator-class CPU, is built on 28nm.
If true, this suggests that AMD is waiting to move to 20nm until that node is more mature, at least for big-core CPUs. The identified process — GF28A — is GlobalFoundries standard bulk silicon node. Other profiles from current and former AMD employees point to Kaveri as being built on 28nm SHP (Super High Performance) which jives with what we’ve heard before.
It’s not clear, at this point, if Kaveri and Carrizo are built on two fundamentally different types of 28nm silicon, or if the different codenames reflect subtle changes. If the 65W target for Carrizo is accurate, it’s possible that AMD is moving to a different node that emphasizes lower power and higher efficiency.
Speaking of roadmaps, however, there’s another leaked variant that hit around Christmas, snapped from a presentation slide.
This Opteron roadmap again puts new Toronto APUs based on Excavator cores and DDR3/DDR4 controllers on the map for 2015, with and without a functional on-board APU. Unfortunately for those hoping that AMD would return to shipping 8-core parts, it also implies that the company is finished with larger CPU configurations — at least for now.
The other interesting bit of information comes from a different engineer who apparently worked on an altogether different project:
There are three possibilities for this: He could be referring to console hardware, to AMD’s GPU technology (possibly in connection with bringing Hawaii back over to GF in preparation for Carrizo), or it could be a reference to Beema/Mullins coming back to GlobalFoundries. Given the difficulty of a 28nm port and the timing, we’re inclined to think this is work being done on AMD’s next-generation graphics core for Carrizo’s hardware — that’s what fits the timeline and makes the most sense.
As for that 28nm-20nm port of an existing design, our best bet there is that this is the ARM core. AMD has no reason to port Carrizo to 20nm, but the Cortex-A57 is expected to shine on that process node.