AMDs new CPUs are more efficient than the last ones - even original would be a big step forward compared to jaguar so thats a good thing.
On PS4/XB1 you had CPU/GPU all on one big chip. Big chips are relatively expensive compared to small ones, as they're all cut from a big round wafer. any defects on that wafer mean the chip can fail and can't be used - increasing the cost of the good ones. The bigger the chip, the more likely a defect will hit it, and the larger the waste (and therefore cost).
So keeping chips as small as possible is more efficient and cheaper.
This new approach from AMD looks to potentially separate CPU and GPU into 'chiplets'. These are then combined onto a common package so they can behave to the overall system like they're one thing, but you're building them out of blocks rather than as one unit. By separating the CPU and GPU into 'chiplets' potentially each one can be smaller and cheaper. Or possibly allows a larger GPU as it doesn't need to share space with the CPU (but this depends on the design of the package and what limits it might have)
Further efficiencies come from having the 'boring stuff' (IO handling etc) outside of those chiplets on the main package. This is done at a larger process (14nm) so should be cheap and mature. And again leaves more of the 7nm chiplet area to be dedicated to the good stuff - actual CPU and GPU processing units, not wasting area on housekeeping.
Very good description. AMD currently has the "uncore" (I/O) @ 12nm using Global Foundries.
I'd add that DARPA has paid AMD to develop Chiplet building blocks for their projects by 2020-2021. The US Department of Energy has paid AMD to develop a Exa-scale computer that will be built with 10,000 APUs that are assembled from GPU and CPU chiplets without video I/O.
The crux of the
DARPA program is to develop a new technological framework in which
different functionalities and blocks of intellectual property—among them data storage, computation, signal processing, and managing the form and flow of data—
can be segregated into small chiplets, which then can be mixed, matched, and combined onto an interposer, somewhat like joining the pieces of a jigsaw puzzle. Conceivably an entire conventional circuit board with a variety of different but full-sized chips could be shrunk down onto a much smaller interposer hosting a huddle of yet far smaller chiplets.
AMD is in the running for
Department Of Energy Invests $258 Million To Build An Exascale Supercomputer
On Thursday, the United States Department of Energy's Exascale Computing Project announced it was awarding six companies --
AMD, Cray, HPE, IBM, Intel and Nvidia -- $258 million to research building the nation's first exascale supercomputer.
An exascale supercomputer would be capable of computing 1 million trillion floating-point operations per second.
The funding will be doled out over a three-year period. The Obama Administration made the first commitment to an exascale supercomputer in 2015. This new set of investments
accelerates the time frame from 2023 to 2021.