witeken
Diamond Member
- Dec 25, 2013
- 3,899
- 193
- 106
Everybody knows that ARM is an inherently more efficient ISA than x86. Don't take my word. Microprocessor design veterans like Jim Keller are of that view
http://techreport.com/review/26418/amd-reveals-k12-new-arm-and-x86-cores-are-coming
Jim Keller was very complimentary about the ARMv8 ISA in his talk, saying it has more registers and "a proper three-operand instruction set." He noted that ARMv8 doesn't require the same instruction decoding hardware as an x86 processor, leaving more room to concentrate on performance. Keller even outright said that "the way we built ARM is a little different from x86" because it "has a bigger engine." I take that to mean AMD's ARM-compatible microarchitecture is somewhat wider than its sister, x86-compatible core. We'll have to see how that difference translates into performance in the long run.
http://www.youtube.com/watch?v=eTVnxaXCLg0
see video at 54:50 to 55:20
This is plain false. ARM isn't inherently more efficient. In modern microprocessor designs, for performance requiring markets, that use multiple watts of energy, ISA doesn't matter at all. This is a pure fact -- supported by scientific research. x86 might not be the best ISA of all time, ARM surely isn't either, and x86 is good enough (Intel tried with Itanium, and they failed).
Intel's also catching up quickly in that area.btw Intel's process lead has not helped them in the mobile market where baseband integration is key. The ARMv8 cores are being actively designed into mobile SOCs.
http://www.anandtech.com/show/7925/...bit-socs-with-lte-category-67-support-in-2015
Of course I realize that those companies don't use off-the-shelf ARM IP, but developing your own cores isn't a free lunch.Qualcomm has standard ARMv8 64 bit CortexA57/A53 cores based top to bottom product stack with integrated baseband in 2015. Qualcomm continues to work on their custom ARMv8 core and will launch in 2016 at 16/14 FINFET.
Competing with Qualcomm is easier said than done. Also you don't realize that the top ARM partners are licensees who design custom cores to differentiate their products - Apple Cyclone, Qualcomm, AMD (K12), Nvidia (Denver), Marvell. That kind of effort is not easy for companies like Mediatek who do not have the resources to design custom cores.
Intel's node naming is misleading.
http://www.electronicsweekly.com/mannerisms/general/the-intel-nanometre-2013-02/
Intel 22nm has no significant transistor density advantage over the foundries 28nm though it does have performance advantage due to FINFET. Intel's 22nm and TSMC/Samsung/GF 28nm all use single pattern immersion lithography.
Intel's 14nm FINFET and TSMC 16FF+ and Samsung 14 FINFET are all of similar transistor density with a marginal lead to Intel. They use dual pattern immersion litho with a M1 metal pitch of 64 nm. TSMC 16FF+ has a 15% area scaling over TSMC 20nm planar. Samsung 14 FINFET has a similar 15% area scaling over 20nm planar designs.
https://markets.jpmorgan.com/research/email/-kjegkq4/GPS-1336259-0
page 4
http://globalfoundries.com/docs/def...dries-14nm-collaboration---final.pdf?sfvrsn=2
If Intel 22nm FINFET Baytrail is competing with 28nm planar tech based mobile SOCs from Qualcomm then you can figure out for yourself how things are going to be with no process node advantage in 2016.
Lol. Intel's node naming misleading? I advice that you read this very informative article: The Status of Moore's Law: It's Complicated.
I'd call Intel's node naming the least misleading. TSMC, Samsung and GlobalFoundries will all switch from 20nm planar to FinFET and call that transition a new node; from 20nm to 16/14nm. Because, they say, going to FinFET will have similar benefits as a normal node shrink. It hasn't. There isn't any density improvement, except for the updated FinFET Plus that will come later, but I don't call 15% a node shrink.
By this flawed logic, Intel would now have been at a 7nm node, soon 5nm (aka 14nm). Truth is, Moore's Law, which is what those transistor sizes are referring to, doesn't tell anything about performance or power, which is why those foundries call it 14/16nm instead of 20nm FinFET; Dennard scaling does.
So in my opinion, it is just plain wrong of those companies to now completely destroy node names. And why? Because marketing; to fool people (I'd see, people like you, but I don't want to sound offensive). In a few years, they will launch 10nm, and someone who didn't inform himself with articles like those mentioned above, will think they caught up to Intel and its lead, which Intel now has for a long time, has vanished. Unless TSMC proves us wrong, which I strongly doubt, this will not be the case at all. Instead, Intel will even have increased its process lead. But people might not know because the number TSMC has given its node says 10 -- same as Intel.
(In easier words: if you buy a 20nm fab, you upgrade the transistor to FinFET, then buy a new, second fab, which node do you have?)