2023+ numbers;
- 22FDX(+, also) NTO (Pre-AI) = $25M design cost + ~$26M verification cost
Actual reason it isn't 22FDX => Not enough density or performance
- 12LP+ NTO (Pre-AI) = $51M design cost + ~$36M verification cost
Actual reason it isn't 12LP+ => GlobalFoundries Digital Solutions group dropped FinFETs for new products. This in-turn killed 12LP+ Single-fin library, for extra density on top of 7.5T Dense Performance.
- 12FDX NTO (Post-AI::OpenROAD.ai and Synposys.ai) = Near-$25M design cost + Near-$25M verification cost
Couple reasons that have been leaked:
- Small track height w/ high mobility
- High Density SRAMs are faster, lower power, lower area than High Currrent SRAMs from 14LPP<->12LP+. Allowing for a full logic+memory shrink, whereas FF-line only touched logic.
- Low end wafer cost, with use of AI.
22FDX = Gen1.1 || 2018 "Mass Production"
22FDX+ = Gen2 (+25% perf) || +3 yrs = 2021 "Mass Production"
12FDX = Gen3 (+25% perf) || +3 yrs = 2024 "Mass Production"
22FDX example before slow-down in 2018:
*risk production will be commencing during 2H 2016
*while mass production of 22FDX chips will begin during 2H 2017.
GlobalFoundries will need to announce 12FDX risk this year for 12FDX mass production next year. Everything so far matches with the comment in 2H 2020: "Globalfoundries expects to be in production on 12FDX in 2023 or 2024."
AMD in the December 23, 2021 WSA added a third 2023 advance payment and the 2025 advance payment. Which is the likely range for the new tapeouts from GlobalFoundries provided by Future Tapeouts, etc.
Extra collection of data:
AMD's WSA has decreasing order size up till the third advancement payment and it increases per each advance payment up to 2025. So, the data means they are going to be buying something at GlobalFoundries in HVM to offset the decreasing sales of 14LPP/12LP devices. This Wafer Supply Agreement is pre-merger with Xillinx.