The race to 10nm FinFET production is still tight with TSMC expected to tape out their first parts towards the end of the year and Samsung today revealing a similar time line according to The Inquirer. Samsung has also confirmed they will be starting construction on a new plant in South Korea in 2017, which is a good move for the company considering their loss of the chip contract for the new iPhone to TSMC. With Samsung going almost full out on their 14nm FinFET lines for the Galaxy S6 and Galaxy S6 Edge Apple had concerns that Samsung would not be able to keep up with demand and unfortunately GLOBALFOUNDRIES could not take advantage either as their yields are, to put it politely, lacking.
"SAMSUNG HAS REVEALED that it will soon begin production of its 10nm FinFET node, and that the chip will be in full production by the end of 2016."
Here is some more Tech News from around the web:
- Boffins silently track train commuters without tripping Android checks @ The Register
- Exploit Kit Delivers Pharming Attacks Against SOHO Routers @ Slashdot
- Mozilla finds a way to tunnel Firefox into iOS @ The Register
- UK government has ended Windows XP support payments @ The Inquirer
- Hisense launches new 4K ULED TVs @ DigiTimes
- Tech ARP 2015 Mega Giveaway : Mi In-Ear Headphones
Well GlobalFounderies(GF) at
Well GlobalFounderies(GF) at least can look forward to the power8/power9 production from both IBM, and any of the OpenPower licensees may just give GF some more business, Samsung could be getting some licensed Power8 action, the Chinese are going to be big OpenPower customers. Both Samsung and GF have been part of the IBM fabrication/other standards technology/IP sharing consortium for some years now, and GF getting a License for Samsung’s 14nm process just smacks of IBM’s involvement in making sure that IBM has more than one fab partner to rely on regardless of the fact that GF has an exclusive contract for IBM’s fab needs. OpenPower licensees will need plenty of fab space so expect not just GF to prosper after the IBM exclusivity contract expires, and as for the OpenPower Licensees they are not bound to any FAB partner exclusivity, and IBM will not let GF screw up it’s delivery schedule for CPU parts. The Power9 will be arriving(1) so expect both of IBM’s fab technology sharing partners(Samsung, GF, and maybe others) to benefit.
Apple appears to go with the fab partner that has its yields sufficient for Apple’s latest designs, so expect Apple to move from foundry to foundry, and often times second source its supply chain. GF yields for parts other than IBM may not be there at the right time to compete for Apple’s business, and TSMC is still in the race but Samsung is in at just at the right time to have it’s 14nm ducks in order, with no guarantees of Apple’s future business. GF does have a 22nm process, via the IBM fab business contract, at least tuned for power8 production, so maybe the current OpenPower customers will work from that process, until power9 come along with whatever process node is ready for it. AMD at least can look forward to both Samsung and GF having an equivalent 14nm process for its Zen based SKUs, so having more than one fab able to supply capacity will not put AMD behind a rock and a hard place should AMD need extra fab capacity. Samsung has plenty of cash to add capacity being such a large conglomerate, GF is just a foundry, Samsung is damn nearly the entire Korean economy.
(1) “OpenPower Lays Out HPC Roadmap”
I don’t think we will ever
I don’t think we will ever see 4nm finFET. commercially speaking.
The economics of going
The economics of going smaller will trump the laws of physics that only allow for gong as small a physically possible. It’s time for more stacking of dies and engineering the transistors to waste less power and produce less heat. There is a lot of materials engineering to be done once the process nodes shrinks on silicon become non economically viable. The low hanging fruit of process node shrinks and easy performance gains have been harvested, so maybe other materials that can be clocked higher and not waste power as heat. Actually redesigning of the gate geometry has just begun with finfet, and the gates may be extended completely around in newer transistor geometries and the logic may go in completely in the vertical Z dimension, there will be logic circuits laid out in a 3d lattice, or other 3d geometries.
Expect RAM memory to move closer to whatever is doing the processing, and HBM is just the beginning, old fashioned RAM not tied directly into the processor via ultra wide data buses will become like a second tier memory store, more for buffering data to a non-volatile flash, as the New NVDRAM standards allows for Flash memory on RAM modules. Hopefully, at least on laptop/mobile systems, SOCs/APU will get at least the option for 16GB OF HBM, and some form of fast PCI 4.O+, or other specialized data lanes to second tier RAM, and Optical interconnects from the Interposer module to the off module memory would be the way to go, so as not to take away the mainboard space savings and complexity reductions that having HBM will allow. HBM is going to be the one thing that will really get SOCs/APUs some energy efficiency in the memory subsystems, and I expect that AMD will be taking their Zen to market with at least some consumer APUs for laptops designed around HBM, and that will be the thing that gets AMD some design wins that will amount to something. OEMs will sure like the savings on mainboard’s PCBs that do not have to have near as many layers for memory traces, Intel better be thinking the same thing because OEM’s PCB costs are not small, especially in engineering, and fabrication of PCBs.
The fab process shrinks are running out of available R&D resources, and economies of scale are not there for any one manufacturer to be able to even reach 4nm, but the entire Mainboard and memory subsystems on PC/Laptops are dew for some attention, and HBM will eventually lead to a complete redesign of the standards.
I see HBM taking over for most computing, with superfast narrow Off module interconnects taking over the of moving data from the interposer Module(CPU, CPU, HBM) to the off module second tier RAM, and storage subsystems, and with all the complexity of the systems main memory moving to the module, some form of super high speed electronic, or optical interconnect would be the way to go in moving data around the motherboard. Future motherboards may look like Islands of processors/other units, with 2 or more optical fibers connecting these islands of interposer modules, the Interposer modules hosting various functional blocks for processing, storage, and routing on and off the motherboard. Motherboards will have only power traces, and optical fibers/wave guides, everything else will be on the interposer modules, including the optical to electrical converters.