Alright, I've been convinced Monet (at least a Zen 3 small die) isn't a thing.
What replaces Dali then? Any speculation on this?
Dali is replaced by Mendocino:
Rumored specs point towards Consumer Van Gogh or lower:
6nm, 4-core Zen2, 8-compute unit RDNA2, 128-bit LPDDR5(4x32), FT6 socket
Mendocino APU(TSMC6FF) vs Monet APU (GF12LP+) => Direct Competition with each other.
The day I see AMD use anything FDX I will owe you an apology.
At this point, you don't...
AMD only has three nodes to pick from:
22FDX -> 22FDX+
12LP+
12FDX
22FDX => lowest design cost + lowest production cost + lowest density
12LP+ => highest design cost + highest production cost + middle density
12FDX => middle design cost + middle production cost + highest density
------
AMD's Zen has moved on to 7nm and below. So, Zen(x) returning to AMD would only be for very small core counts: ~2-cores.
However, these nodes are still within the range of the low-cost/low-power cores since they were never ported down from 28nm.
If 12FDX => Old Family 16h cores, but are models we never saw before. (Margay+Vega+Infinity Fabric)
If 22FDX => New Family cores. (New CPU, New GPU, New Fabric), this would be the second product for 12FDX.
If 12LP+ => GlobalFoundries will kick AMD out when #1 in wafer volume and #1 in revenue moves from 14LPP to 12FDX.
22FDX++/12FDX with SiGe PFET and Tensile Si NFET is probably what AMD is waiting for...
28HP -> 28HPP -> 28SHP(AMD 1st Kaveri) -> 28A(AMD 1st Bhavani), AMD at GlobalFoundries given 28nm waits for the highest performing implementation. There is also a case for multi-fab within GlobalFoundries. A node isn't successful at GloFo unless it is run across multiple modules+fabs.
Which is why I focus on 22FDX, but they could always do a 32nm and aggressively push a 12FDX product out. Since, it uses the same transistors as 22FDX at higher density.
12FDX:
Opteron X successor => Datacenter -> Return to dense computing (Ex: Turing Pi (1/2), SeaMicro-esque, OmniCluster SlotServer1000, etc)
Sempron successor => Personal computing -> $36 and below. (28BLK in 2014 to 12FDX in 2024 is more than a half-price drop, same ASP -> higher profits)
G-series successor => Embedded and other emerging markets -> G1000 series under R1000*ded*/R3000*Mendocino* series
- One die favors 12FDX (No seperation of CPU and APU), also all CPUs with Raphael and after have integrated GFX with cIOD3.
- 22FDX/12FDX having lower cost than 28nm/14nm allows for Import&Custom busting: $36 APU + $36(1st tax/tariff) + $36(2nd tax/tariff) + $36(etc^x-2). => More recent laws implemented 2018+. Example with Cezanne: $350 APU + $700 Tariff => $1050 APU in Latin America.
GlobalFoundries also sort of dropped the lead times:
30,000 wafer per month for 14LPP/12LP/12LP+, if it all swapped over to 12FDX, it would be close to 60,000 wafer per month.
14LPP/12LP/12LP+ only at Malta -> AMD's WSA probably be moved to bidding for volume of wafers.
12FDX only at Dresden and Malta -> Revival/successor of 28nm two foundries, AMD probably won't need to bid for capacity. Capacity bid is to get a profit from a money loss at GlobalFoundries: Fotonix/FDX has a profit at GF, while Fins are run at a loss.
On 12FDX there is Ocean12 which had a duration of 45 months: 4th Month of 2018 plus 45 months is 1st Month of 2022. Which has been extended a bit:
12FDX 2018~2019 Gen1 Wafers:
- 7nm Performance/Power (Only 2018 Leti-involved performance boosters in 12FDX PDK, future saved for 12FDX+)
- Low volume and high variation from SOITEC
- 2017
12FDX 2023-2024 Gen2 Wafers:
- 5nm Performance/Power (2018-2023 Leti-involved performance boosters in 12FDX PDK)
- Higher volume and lower variation from SOITEC
----
Why low-cost/low-power focus at GlobalFoundries:
Value/Mainstream/Performance "High Performance aka Zen" = ~60% of production
Budget "Low Power aka FDX cores" = ~40% of production
If GlobalFoundries 12LP+ is used with Monet, it would be competing in the smallest section with Mendocino. Of the 60% going to Zen in this case, only 20% of Zen is going towards Value. Where if AMD restarted low-power/low-cost it would take 100% of Budget.
TSMC gets 100% of AMD's focus with high performance and GlobalFoundries gets 100% of AMD's focus with low power, etc.
22FDX competition; 2c/4c+weak GPU -> Same cores for 2c/4c:: Cortex A72-ARM and RISC-V
12FDX competition; 8c+medium weak GPU to no GPU -> ARM is big.Little 2xA76+6xA55 or custom 8x(FTC663), RISC-V is 8x A75, 8x x86-64 is not much better than Isaiah II or non-existent.
If just a straight Jaguar to 12FDX to get a product out with GlobalFoundries once it goes in production:
Phytium D2000 => 132.08 mm2(Given by Phytium)
Zhaoxin ZXE => 144.71 mm2(Pixel measurement)
Succeeds Pollock (6W-64-bit) => 149 mm2:
It is safe bet for me to guess (8 core + 4 compute + 64-bit DDR5) will be around ~107 mm2.
I wouldn't suggest completely static IPC, since Llano increased Retire/Reservation/MemoryQ/etc. Not much is known about the 2016-2018, 2020+ ULP CPU/GPU/Fabric/Cache. Which can be like Bobcat<->Jaguar, where it is more portable in regards to time to market.
Also have to deal with speed from Automated-ML PNR with improvements from...
Bobcat:
"Except for a few custom memory arrays, the core is fully synthesized, so it can be ported quickly and efficiently to different process technologies."
Jaguar:
"JG is designed with area as a priority. Standard place and route tools are used to construct the core as a single block with 1.25M placed instances."
"To further improve the design efficiency over previous cores, JG uses only 2 unique memory modules, a RAM and a ROM. The BT core had 7 unique macros. Reducing the number of custom macros reduced timing closure iterations, resources needed to design the macros, and enabled easier process porting."
Just so everyone knows for 28nm FDSOI, STMicroelectronics High Perf A9:
"This paper presents the implementation details and silicon results of a 3 GHz dual-core ARM CortexTM-A9 (A9) manufactured in the 28 nm planar Ultra-Thin Box and Body Fully-Depleted CMOS (UTBB FD-SOI) technology. The implementation is based on a fully synthesizable standard design flow."
EU-12FDX covers prior-28A production facility and US-12FDX covers prior-14LPP/12LP production facility. 22FDX covers older 90-nm Chartered production facilities unless Singapore's new modules include 12FDX through 22FDX support. Which is another bullet point for both 22FDX/12FDX and not 12LP+.
Shortened:
22FDX Need-list:
2c or 4c at ~A72 IPC
3 CU with improvements over 28nm 2CU/3CU.
DDR4-3200 to DDR5-4800
12FDX Need-list:
8c at ~A75 IPC
4 CU with improvements over 14nm 3CU.
DDR5-4800 to DDR5-6400
Can't be done at just Malta thus 12LP+ is dead on arrival.
Off-topic:
22FDX Wafer => 12nm tSOI
Next-Gen FD Wafer => Down to 5nm tSOI
22FDX has range of 8nm SOI(2/3rds of tSOI) to 6nm SOI(1/2th of tSOI)
Does that mean 12FDX has a range of 3.33 nm SOI(2/3rds of tSOI) to 2.5 nm SOI(1/2th of tSOI)
Just need secondary help on how big range of DIBL/SS/Mobility improvements are for 12FDX over 22FDX.
www.cnx-software.com
Facebook
Twitter