So it's going to be useful for (relatively niche) embedded stuff, not mass market consumer stuff? As expected.
AMD Embedded G-Series J Family SoCs are designed to meet the processing requirements of a wide range of embedded applications, including:
- Thin Client
- Digital Signage
- Digital Gaming
- Retail Point of Sale
- Industrial/Automation
- Military/Aerospace
- Media & Collaboration
- Smart Camera and more
//AMD recently refreshed its AMD Embedded G-Series J Family of processors, formerly codenamed “Prairie Falcon,” ...
https://community.amd.com/community...sionGUID=a5516c33-847a-a83d-e1d9-890483779d7e
^-- Last modified on Jun 12, 2018 9:02 AM
^-- 2018 refresh
http://linuxgizmos.com/files/amd_3rdgengseries_j_chart.jpg
^-- 2016 numbers
The whole cost advantages of 22FDX, low cost and low risk, are not really there for 12FDX so I don't think the prospect is so good.
So, 22FDX is either done at AMD or since the pivot moved all tasks to 12FDX.
The extra cost of 12FDX is buffered by the shrink.
22FDX => 125 mm2 * <0.78 => <97.5 mm squared
12FDX => 125 mm2 * ~0.43 => ~53.75 mm squared
For a Stoney Replacement, the 12FDX node builds upon 20LPM/14XM/14LPP nodes.
-> First 12FDX SoC => Cost reduction, power reduction, complexity reduction, and sub-cm2 fast HVM.
XV Watt Range per module is 2W to 18W.
22nm-FD & 20nm-BLK => <1W - <9W for same performance.
12nm-FD => <0.5W - <4.5W for same performance.
The main reason the actual watts would be less than is because of Vt re-selection and track height reduction. 22FD is 8-track and has lower power Vts, 12FD is 7.5-track and has even lower power Vts. There might also be IP redesigns, and new units to add, but ~25W performance should be converged into ~6W power.
Edit: 18W is 3.5 GHz to 3.7 GHz
12FDX at 9W would have the whole 50%(22FDX) + 26%(12FDX) frequency boost at same power. There is also 12FDX superior body-biasing, allowing for dual-BB(FBB+RBB). At idle all the transistors can reverse body bias, at active all the transistors can forward body bias. Whereas, in 22FDX it is optimized for RVT/HVT for reverse, and LVT/SLVT for forward.
12FDX allows for a more simplified AVFBS; Adaptive voltage-frequency-bias scaling, than 22FDX. Body-biasing would allow for a superior embedded with reliability-focused body bias, and a superior consumer with performance-focused body bias.
;End edit
The two entry-products for 2020;
N7+ => 2x Zen3 + 3x GPU @ $60 price; sustained ASP and premium.
12FDX => 2x NCx + 3x GPU @ $10 price; reduced ASP and budget.
Going up;
N7+ => >4x Zen3 + <16x GPU
12FDX => ~4x NCx + <8x GPU
N7+ => >16 Zen3
12FDX => ~8x NCx
N7+ => 24 CUs
12FDX => ~6 CUs
I wonder if they would use 22FDX for the active interposer, once they get to that.
The FDX platform is aimed at 3DVLSI.
https://www.semanticscholar.org/pap...madi/8d52a706230a6056844b663a1d4cb1dfb97eb7fa
https://ai2-s2-public.s3.amazonaws....a6056844b663a1d4cb1dfb97eb7fa/2-Figure1-1.png
Bottom two dies would be 12FDX, four top dies would be 22FDX. If it was a GloblFoundries slide.
The only 2.5D currently available for the FDX platform is Organic and MCM. While only the FinFET platform has Interposers, TSVs, Si Photonics.