Whereas 14nm and 12nm are installed, running, and profitable.
No, GlobalFoundries 14nm/12nm FinFETs is sold at a loss.
If AMD will guarantee enough wafer purchases to keep those production lines full until 2025, then they will keep running. That is pure profit for GloFo.
Pure loss for GlobalFoundries for the one major reason of 45nm SOI being so big as it is at Fab8. SOI is profitable where FinFETs are not.
The adoption of 14LPP was meant to attract customers from Samsung. Which went nowhere since Samsung did 14LPC/14LPU/11LPP and GlobalFoundries did 12LP/12LP+. As is 14LPP/12LP/12LP+ has the lowest figures of merit.
Fab8 does not have anything competing against Intel Oregon's 22FFL or Samsung's Austin 17LPV/18FDS. It also doesn't provide better options than 14GP-Intel and 11LPP-Samsung.
There is also the removal of exclusivity for 14nm/12nm FinFETs. Which means AMD does not have to use GlobalFoundries for N-1 FinFETs products. The outcome is that AMD is making a move away from GlobalFoundries.
TSMC 16nm/12nm is the new home for Ryzen/EPYC for N-1 going forward. Not AMD committing more money for N-1 at GlobalFoundries.
AMD will move to the tune of GlobalFoundries or be trimmed out. Nothing is stopping GlobalFoundries in making a better business investment. Remember the pivot cut off five customers who were pathed to 7nm. Which AMD was #1 customer of at first.
GlobalFoundries moves Malta's 32nm/28nm capacity to Dresden.
GlobalFoundries runs purely 14nm/12nm at Malta.
First year = Negative net income || Dresden&Singapore Positive Revenue / Malta Negative Revenue
Second year = Negative net income || Dresden&Singapore Positive Revenue / Malta Negative Revenue
Third year = Negative net income || Dresden&Singapore Positive Revenue / Malta Negative Revenue
Fourth year = Negative net income || Dresden&Singapore Positive Revenue / Malta Negative Revenue
Fifth year = Negative net income || Dresden&Singapore Positive Revenue / Malta Negative Revenue
GlobalFoundries starts "GlobalFoundries will transfer its ongoing operations from Fab 10, as its East Fishkill fab is called, to its other facilities, including the massive Fab 8 plant in Malta."
Sixth year = Lower negative net income || Dresden&Singapore Positive Revenue / Malta Lower Negative Revenue
Seventh year = Brief positive net income || Dresden&Singapore Positive Revenue / Malta Brief Positive Revenue (45nm SOI surpassed 14nm/12nm FinFET wafer starts)
Then, there is the FDX node family also taking space of Fab8 eventually:
In turn converting Malta from big loss-zilla to big gains-kong.
For 2023:
12FDX/22FDX same tools used --{ 14LPP/12LP/12LP+ }-- 45SOI same cleanroom space
With very little to no capacity planned for FinFETs as they get squeezed by demand from tens of billions multi-customer rather than hundred of millions single-customer. In this case, the wants of the many outweigh the wants of the one. However, I am not the one thinking AMD's only purchase going forward is only 14nm/12nm FinFETs at GlobalFoundries.
We can use UMC's revenue as a guide for GlobalFoundries profitability, since their footprint is the same.
Negative revenue to income contribution by 14nm/12nm.
Positive revenue to income contribution by 28nm/22nm.
Positive revenue to income contribution by 55nm/45nm/40nm.
etc.
Jason Wang of UMC => "In addition, we expect the adoption rate of our 22nm technologies will continue to gain traction" - Q2 2021
"Revenue from 28-nanometer technologies continue to rise while business engagement in 22-nanometer have led to a growing number of customer tape-outs across wireless, display and IoT markets, further diversifying our product pipeline." - Q3 2021
Do to 28nm demand the price went from $1600 per wafer to $2300 per wafer.
2021-05-11 "The 28nm process price per wafer is approximately US$1,800, an increase of nearly 13% from the US$1,600 in the second quarter. In addition, industry insiders said that UMC will raise prices again in the first quarter of next year. By then, the price of 28nm process wafers will soar to US$2,300 per piece, an increase of nearly 28% from US$1,800."
2021-12-25 "For the same 28nm process technology, TSMC's previous quotation was US$2400-2800; for comparison, UMC's quotation was US$2100-2200, GlobalFoundries' quotation was US$2100-2300, and SMIC's quotation was US$1800-1900. Based on the current TSMC 28nm process quotation, it is estimated that the highest price is about US$2,800. After a 20% increase, it will be higher than UMC's 28nm process technology in January 2022 to nearly US$3,000."
GloFo's 28nm is 2100-2300 thus 1.95x 14nm193i = $4100-4485
The several customers for FinFETs(2020 Annual Report) at UMC have yet to appear. However, the more than 10 customers(2020 Annual Report) for 22nm have definitely appeared.
UMC = $627 million net income, Q3'21
GloFo = $5 million net income, Q3'21
This is reflected by GlobalFoundries getting design prototype wins for 14LPP/12LP/12LP+ but the design production is done at Samsung or TSMC instead. In this case, AMD is transferring cIOD/sIOD/Raven2/Raven/Summit/etc dies over from GloFo to TSMC as soon as possible.
There is no further adoption of 14nm/12nm FinFETs to Dresden or Singapore. However, there is adoption indications of 22FDX at Malta and Singapore. Thus reducing 22FDX load of, NA+LA markets with Malta and GC+APJ markets with Singapore. In turn giving more 22FDX to each major region more than one Fab in one region can.
The Singapore option is the most important since the extremely low-cost Turion/Sempron processors were fabbed over there.
Going further some analyst groups in 2021 have noted 28ULP(28SHP/28A rebrand) and 22FDX is getting prepped at Fab8.
The 28SHP/A to 28ULP renaming follows UMC's weird DDC node names 28HPM -> 28HPC -> 28HPCu+ -> 22ULP.
AMD's May 2015 Low-cost cancellation is dependent on increased costs from 20nm -> 14nm -> 7nm -> 5nm. While, Low-cost by AMD coming back is based off 28nm -> 22nm FDSOI -> 12nm FDSOI which have decreased costs going forward.
Margin's decrease 20nm -> 14nm -> 7nm -> 5nm; Low-cost is unsustainable
Margin's increase 28nm -> 22nm -> 12nm -> 6nm (7FD CEA-Leti/IMEC/GF); Low-cost is sustainable
AMD May 2015 = Low-cost is dropped
GloFo July 2015 = 22FDX is announced
AMD 1H2016-2H2018 = Ultra-Low-Power Architecture for CPU&GPU
GloFo September 2016 = 12FDX is announced
CEA-Leti and #1 Foundry Partner 2020/2021 = 7FD(44CPP/36Mx) is announced in the works
Dali Mobile = January 2020, Dali Desktop = July 2020, Pollock Mobile = July 2020
28nm EOL = January 2021 LTB(Last Purchase), July 2021 LTS(Last Recieved)
Mendocino Mobile = 2022, Mendocino Desktop = 2022 (killing off demand of 3000G(OEM)/3050GE(OEM))
Raven2/Dali >$50 -> Mendocino >$100
Stoney/Carrizo-L >$25 -> 22FDX >$25
Dali/Pollock >$50 (100% performance at Iso-freq/150 mm2) to 22FDX >$25(90% performance at Iso-Freq/90 mm2)
Stoney/Carrizo-L >$25 (~70% performance at Iso-freq/107-125 mm2) to above.
CPU/GPU is not as critical when the most important aspect is the Multimedia core.
14LPP/12LP+ Multimedia core
Reducing leakage by Vt = RVT -> HVT -> sHVT -> eHVT (four options; HVT onwards costs process steps)
Reducing leakage by Gate = 14nm Standard Gate -> 16nm Wimpy Gate (two options; wimpy costs masks)
22FDX/22FDX+ Multimedia core
Reducing leakage by Vt&Vbb = RVT -> RVT w/ RBB -> HVT -> HVT w/ RBB (two options with boosters; no added process steps)
Reducing leakage by Gate = 20nm -> 24nm -> 28nm -> 32nm -> 36nm options (five options, no added masks)
Siginifigantly lower power thus longer battery life when doing the ole Youtube/Netflix/Disney+/Amazon Prime/etc binge shuffle.
External factor: GlobalFoundries has increased demand for 45nm SOI and 22nm SOI platforms. In turn, GlobalFoundries supply for 14nm platform is reduced as it isn't in demand nor is it contributing to income.
Internal factor: AMD has desire to re-enter low-cost via ultra-low-power CPU(AMD64/RV64)/GPU(Post-GCN)/SoC(Multimedia, Crypto-Security, RF, Integrated I/O, Vision, Semi-custom IP).
Combined factor: AMD/GF going forward partnering up on 22FDX onwards.
Proof of risk-aversion:
1. AMD removing exclusivity of 14nm/12nm node. (GloFo revenue for 14nm/12nm is less secure)
2. AMD getting lower price of 14nm/12nm node. (GloFo revenue contributes less to income)
With a return of micro-server targets because of ULP architecture and lower costing node.
I have seen the above but with fanless 4x 22FDX CPU/GPU/APUs with an eight slot cluster motherboard.
With Topaz target allowing for equivalent of: 6 CUs * 4 dies * 8 cards => 192 CUs at ULP arch/clocks.
Lower cost of entry => Higher volume of sales to consumers
Reduced cost of fabrication => Higher margin of profit to AMD
Pre-2015:
HP and LP are on the same node with low-cost option having reduced margin.
| Node | HiPerf - HiCost | LowPower - LowCost |
| 28nm | Viable | Viable |
| 20nm | Viable | Non-Viable |
| 14nm | Viable | Non-Viable |
| 7nm | Viable | Non-Viable |
2016-2018 Post-2019:
Splitting HiPerf and LowPower to separate nodes
| High Performance Architecture + Increasing ASP, Increased Margin | Ultra Low Power Architecture + Recession Tolerant, Increased Margin |
| HP = 14nm/12nm | ULP = 22FDX || 2022+ :: 12FDX RiskProd |
| HP = 6nm/5nm || 2022 | ULP = 12FDX || ~4 year cadence = 2026+ :: 6FDX RiskProd |
| HP = 4nm/3nm || 2024 | ULP = 6FDX || ~4 year cadence = 2030+ |
cIOD/sIOD 12nm is easily replaced by bin harvested cIOD/sIOD 6nm. There is also second source on 16nm/12nm at TSMC becoming first-source going forward.
Pollock 3015e = July 2020 ~4 year cadence => July 2024
28nm end of manufacturing = July 2021 ~2 year substitute => July 2023
A6-9220C/A4-9120C = January 2019 ~4 year cadence => January 2023
22FDX insert is January 2023 through July 2024.
2017/2018/2019 = A55/A65/E1 [A65/E1/Helios was originally meant for 2015-20nm]
2021+ = A510, no A610/E2 OoO SMT core yet.
2025+ = A530 Hypothetical, A630/E3 Hypothetical
^-- Target of ULP AMD64/RV64
AMD ~2.7 million units for HP cores, whereas previous low-cost cores shipped ~6 million units per quarter.
Without the rush to get rid of margins to the next node by the prior low-cost core:
1.
"Lynx Family" 40nm Bobcat(2011)/28nm Bobcat(2012)
"Panthera Family" 28nm Jaguar(2013)/20nm Leopard(2015) || "Puma Family" Puma = GF28 and Catamount = GF20
"Leopardus Family" 16nm/14nm Oncilla/Margay(2017)
2.
22FDX ULP1 (2022+)
12FDX ULP2 (2026+)
6FDX ULP3 (2030+)
HP core - Increased costs from FinFET, increased TDP from increased transistors leading to higher Pstat/Pdyn/Ptotal, higher cost of entry: low customer count.
ULP core - Lower costs from FDSOI, decreased TDP from reduced transistor count lower Pstat/Pdyn/Ptotal, lower cost of entry: high customer count.
But more to the point- 7nm required a lot more R&D and investment in equipment. Whereas 14nm and 12nm are installed, running, and profitable. If AMD will guarantee enough wafer purchases to keep those production lines full until 2025, then they will keep running. That is pure profit for GloFo.
www.notebookcheck.net
www.dolphin-ic.com
www.dolphin-ic.com
Facebook
Twitter