Over the previous few years, firms like Samsung, TSMC, and GlobalFoundries have all jockeyed for the pole place within the extremely aggressive (and very costly) foundry enterprise. Even Intel, which used to march to the beat of its personal drum, has gotten in on the motion. As the price of every new node has risen, the should be first has elevated as effectively. The primary firm to hit a brand new node tends to safe most of the most profitable contracts. For a lot of the previous 20 years, the pure-play foundry world was dominated by TSMC, however Samsung beat its Taiwanese rival to 14nm and made a good bit of money within the course of.
Anandtech has put collectively a comprehensive discussion of how every foundry is approaching the following few generations of Moore’s Law, the applied sciences they intend to deploy, and which firms will debut which nodes (and on what time frames). The entire piece is completely value a learn, particularly if you wish to know the specifics of how Samsung’s low energy course of is predicted to match with TSMC’s. It’s a wonderful deep dive. We’re going to concentrate on the highlights, as proven within the chart under:
Let’s take ’em so as. GlobalFoundries will follow 14LPP for the following 12 months, however expects to start excessive quantity manufacturing on 7nm by the top of 2018. The start of HVM is not the identical factor as transport shopper merchandise; it will probably take Four-7 months for gadgets to really hit the market as soon as manufacturing spins up. 7nm DUV (standard 193nm lithography) ought to start within the again half of 2018, and GF desires to have EUV and 7nm able to go by 2019.
We’ve mentioned Intel’s foundry plans earlier than (the slideshow under has further particulars), and the way its 10nm know-how is meaningfully totally different from the 10nm constructed by Samsung and TSMC. Intel will start rolling out 10nm in cell this yr, however desktop and HEDT are anticipated to stay on 14nm, with a third-generation 14nm++ debuting within the again half of this yr and into 2018. Intel hasn’t made any particular statements about EUV, although the corporate has implied it could not introduce the know-how till the 5nm node.
The Samsung-TSMC match-up is an attention-grabbing one. TSMC seems to have extra capability on-tap for 10nm and is planning a fast transition from 10nm in 2017 to 7nm in 2018. Samsung, in distinction, is planning to carry at 10nm for a number of product generations. As nodes have develop into harder to hit, every foundry has adopted varied strategies of defining what constitutes a node shrink. That is most obvious with Intel; its 10nm node is predicted to be equal to 7nm from TSMC.
Samsung seems to be planning a relatively late EUV introduction. Its 8LPP can be an 8nm node constructed with standard immersion lithography, whereas its 7nm node introduces EUV in 2019 – 2020, with the CLN7FF+ node.
Will EUV be prepared?
There are two issues to remember when evaluating these roadmaps. At first, EUV (Excessive UltraViolet lithography) stays a serious confounding variable. TSMC, GlobalFoundries, and Intel all need to introduce EUV, as a result of it gives large potential enhancements over double, triple, and quadruple patterning we see at present. The issue is (and has been) that EUV is each costly to deploy and never at the moment hitting the efficiency targets it wants to succeed in to be a viable substitute for 193nm lithography.
The second factor to remember is we don’t understand how effectively any given firm will execute and we will’t actually examine nodes primarily based solely on their arbitrary numbering anymore. When TSMC was the one sport on the town, it was simple to match their 65nm to 45nm, or 90nm versus 65nm. Now, the scenario is far more chaotic, with a number of revisions of every node and every firm adopting its personal naming scheme.
The truth that Intel, TSMC, and GlobalFoundries have all begun adopting a number of variations of the identical node speaks to how troublesome it’s to proceed to ship constant efficiency will increase. Ten years in the past, the thought of frequently iterating on a single course of node made no sense in any way. Right this moment, it’s develop into frequent place.
I believe (although can not show) it’s not simply the foundry nodes that get tweaked by means of this course of. There’s a multi-step studying course of by which the foundry tweaks its know-how to higher match desired traits of a given class of gadgets, whereas the developer of stated gadgets (Qualcomm, Intel, AMD, and many others) tweaks their design to higher match the method. This was lengthy thought of one in every of Intel’s best strengths, although its determination to exit the pill and smartphone market principally eliminated it from consideration when evaluating with the pure-play foundries.
Once more, I like to recommend giving Anandtech’s full story a glance, significantly should you’re interested in promised enhancements at varied nodes from particular firms.
Now learn: The myths of Moore’s Law