Analysts compare TSMC/Intel/Samsung nodes: Intel still at the top

[witeken]

EETimes: Silicon Lacks Clear Metrics

A lot of confusion, though.

Earlier HVM and superior metrics across the board: I'd say still a good 2 year lead, heuristically.

TSMC has been clear its 16nm process uses its 20nm back-end technology with FinFET transistors layered on top. At its recent symposium in San Jose it said its 7nm node will be 1.63 times denser than its 10nm process, “which is well short of 2x that would result from 0.7x linear scaling in both dimensions, yet the node name was scaled 0.7x,” said Wei of Chipworks.

 

[ShintaiDK]

Only people fooling themselves would think otherwise

The future will be interesting tho, how many are PR nodes and how many delivers. And on what time, what cost and what yields.

 

[Phynaz]

GloFlo is only mentioned in one sentence in a two page article. Goes to show their relevance as a leading edge manufacturer.

How long is AMD tied to them?

 

[frozentundra123456]

Maybe, I dont know enough about transistor layout to say. But seems to me all the talk about this node is better than a competitor's is kind of moot. What counts is the final performance of the products.

I dont really see how Intel can be 2 years ahead and still not make a competitive mobile chip. It will be interesting to see how 14 nm performs on Zen.

I do think whatever lead Intel has is in danger because of slowing to a 3 year cycle instead of 2. For sure they cant keep relying on process node superiority forever, because eventually they will hit the cost/physics wall. Not to mention that each new node seems to be bringing less and less benefit.

 

[Eug]

^^^ Regarding that first comment in the above post, it is interesting that table includes Apple A9, with the die sizes, because it was believed by some that the larger TSMC chip was the superior one over the Samsung one. I'm thinking "superior" in terms of yields and possibly even power usage. Furthermore, it was also believed that TSMC won Apple's contract for A10 partially based on what happened with the dual-sourced A9.

 

[scannall]

It shows that Intel still has a process lead. No surprise. And that their lead has diminished. No surprise there either.

 

[ShintaiDK]

GloFlo is only mentioned in one sentence in a two page article. Goes to show their relevance as a leading edge manufacturer.

How long is AMD tied to them?

Only till may 2nd 2024. D:

Its really a shame they are tied to such a bad foundry. Had it just been TSMC instead it wouldn't matter.

 

[el etro]

Only till may 2nd 2024. D:

Its really a shame they are tied to such a bad foundry. Had it just been TSMC instead it wouldn't matter.

A Wrector's Heritage. He even got a chair at GF BoD with this.

The other one is Bulldozer(ok, Dirk have some fault here, but CON itself started developing under his term).

 

[myocardia]

A Wrector's Heritage.

Wrector Ruinz...I really like that. Consider it officially stolen.

 

[dark zero]

Only till may 2nd 2024. D:

Its really a shame they are tied to such a bad foundry. Had it just been TSMC instead it wouldn't matter.

Getting tied to TSMC would be on an eternal hiatus....

 

[Shehriazad]

I feel like Intels' lead will shrink down to 0 or almost 0 in that regard...because new materials that allow us to go smaller in the single digit area are either not fully developed or SO far from cost effective that it ain't even funny anymore.

 

[Madpacket]

Samsung looks pretty close to Intel all things considered. Didn't AMD and Glofo license Samsungs 14nm tech for Polaris, Vega and Zen?

 

[IntelUser2000]

I'm thinking "superior" in terms of yields and possibly even power usage. Furthermore, it was also believed that TSMC won Apple's contract for A10 partially based on what happened with the dual-sourced A9.

Because of reasons like these its hard to say who's really #1. Based on the simple data, Samsung is better than TSMC, but when some reviews were showing TSMC chips superior in metrics that's more important like power usage, who cares about the 10% bigger chip?

IMO its possible that Intel actually has the worst process. They might have pushed for density unnecessarily to "win" and delays and mess-ups might have caused them to lose everywhere else: Performance, Power Usage and Leakage.

Especially in mobile where TSMC has much more experience. Also because they make completely different chips and in that area it showed meagre improvement, I'd dare say their 14nm sucks.

This isn't an attack on their engineers. The management and those related that possibly due to inflated egos did more than the team can handle is at fault. What if they really did put all their eggs in one basket?

 

[ShintaiDK]

Samsungs 14nm node is based on industrial espionage from TSMC. So I wouldn't credit them anything.

 

[raghu78]

TSMC is fast closing the gap between themselves and Intel. TSMC's 10nm node is scheduled for volume production in Q1 2017 and TSMC 7nm which uses the same BEOL as TSMC 10nm while improving FEOL will go into production in Q1 2018. Apple is driving the ramp of TSMC 10nm and TSMC 7nm nodes. In fact TSMC 10nm node will have few customers just like TSMC 20nm. Xilinx has confirmed they will skip TSMC 10nm and go straight to TSMC 7nm. I think GPU companies like Nvidia and HPC customers will go straight to TSMC 7nm. btw the naming of these process nodes is not comparable across all companies. TSMC 7nm will have similar transistor density as Intel 10nm. TSMC has also confirmed a 7nm High performance node for HPC.

http://www.eetimes.com/document.asp?doc_id=1329217&page_number=2

We could see Intel 10nm and TSMC 7nm products face off in HPC markets in 2019.

 

[Phynaz]

Is TSMC producing big 4Ghz+ parts? What are their yields and costs? The problem with figuring out who is a ahead is it all depends on how you are measuring it.

 

[ShintaiDK]

Yep, Intel is manufactoring 400-600mm2 dies in HIGH volume with HIGH yield. Not to mention the 4.2 and 5.1Ghz parts.

 

[frozentundra123456]

Well, I think Intel has the lead in high performance parts. Perhaps we will know more when Zen comes out how Samsung 14nm performs at high frequencies.

The problem for intel is mobile. I mean phones, and to a certain extent tablets. They really needed a lot more from atom at 14nm as far as power savings and cpu performance. I dont know if it is the process, or the design, or both, but for anything below core they just are not competitive.

 

[ShintaiDK]

Looking at the power and performance metrics on other parts. And taking history into the accounts I am 100% certain its the design. Atom is simply stuck between 2 nails.

And Core M is reaching the point where it can power a phablet, if you wanted to.

Intel just need to kill of the Atom line and make a Celeron M or something from the Core line.

 

[SarahKerrigan]

Is TSMC producing big 4Ghz+ parts? What are their yields and costs? The problem with figuring out who is a ahead is it all depends on how you are measuring it.

SPARC M7 is a specialized 32-core 4.13Ghz server chip on TSMC 20nm. SPARC64 X+ is a 16-core server chip at 3.7GHz on TSMC 28nm.

 

[el etro]

TSMC will have a FPGA/HPC 10nm variant from what i know, so their official HP processes will come back. They also classify their 16FF+ as a high-performance focused node.

 

[Glo.]

The process of Samsung that is dated at 1Q15 is LPE, not LPP.

Also do not forget that GloFo bought not only Fab from IBM, but also engineers, and technology for FD-SOI.

 

[ShintaiDK]

Glofo didn't buy, they got paid to take it. That how a big success that one is

 

[frozentundra123456]

Looking at the power and performance metrics on other parts. And taking history into the accounts I am 100% certain its the design. Atom is simply stuck between 2 nails.

And Core M is reaching the point where it can power a phablet, if you wanted to.

Intel just need to kill of the Atom line and make a Celeron M or something from the Core line.

Well, they are kind of between a rock and a hard place there. I know the ark intel price is not what the OEM pays, but the cheapest core M processor is 281.00, while the price of the 3735F is 18.00 for instance. I am sure the core M doesnt cost 15x as much to make, but I dont know the cost structure and how much margin they would be willing to sacrifice to get core M into phones and more tablets.

I would gladly pay 200.00 for a nice core M tablet, but most are stuck with atom or very expensive like Surface 3 and Surface pro. Doesnt even the most efficient core M still use quite a bit more power than atom though? I mean in real life use, not just based on TDP or SDP or whatever.

Edit: maybe they should just fire everyone who was ever associated with Atom and use the cost savings to lower the price of core M. Just kidding, sort of, but one does have to wonder if Atom is worth the resources devoted to it.

 

[jpiniero]

Core M draws way too much to be used as a phablet. I expect this is Intel's #1 priority with Cannonlake but until it launches you'll just have to wait.

Edit: maybe they should just fire everyone who was ever associated with Atom and use the cost savings to lower the price of core M. Just kidding, sort of, but one does have to wonder if Atom is worth the resources devoted to it.

Part of the problem is that Intel doesn't want to make Atom "too good".