This site may earn affiliate commissions from the links on this page. Terms of apply.

moore's law head

Over the past few years, companies similar Samsung, TSMC, and GlobalFoundries have all jockeyed for the pole position in the highly competitive (and extremely expensive) foundry business concern. Fifty-fifty Intel, which used to march to the beat of its ain drum, has gotten in on the action. As the cost of each new node has risen, the demand to exist starting time has increased too. The offset company to striking a new node tends to secure many of the most lucrative contracts. For about of the past two decades, the pure-play foundry globe was dominated by TSMC, but Samsung trounce its Taiwanese rival to 14nm and made a fair scrap of cash in the process.

Anandtech has put together a comprehensive discussion of how each foundry is approaching the adjacent few generations of Moore's Police force, the technologies they intend to deploy, and which companies volition debut which nodes (and on what fourth dimension frames). The whole piece is absolutely worth a read, especially if you want to know the specifics of how Samsung's low power process is expected to compare with TSMC's. It's an excellent deep dive. We're going to focus on the highlights, equally shown in the chart below:

Let's take 'em in club. GlobalFoundries will stick with 14LPP for the next 12 months, but expects to begin high volume manufacturing on 7nm past the end of 2018. The beginning of HVM is not the same thing as shipping consumer products; it can accept 4-7 months for devices to actually hitting the market in one case manufacturing spins upward. 7nm DUV (conventional 193nm lithography) should begin in the back one-half of 2018, and GF wants to have EUV and 7nm ready to go past 2019.

We've discussed Intel's foundry plans before (the slideshow below has additional details), and how its 10nm applied science is meaningfully dissimilar from the 10nm congenital by Samsung and TSMC. Intel will begin rolling out 10nm in mobile this year, merely desktop and HEDT are expected to remain on 14nm, with a 3rd-generation 14nm++ debuting in the dorsum half of this twelvemonth and into 2018. Intel hasn't made whatsoever specific statements nigh EUV, though the company has implied information technology may not introduce the technology until the 5nm node.

The Samsung-TSMC match-up is an interesting one. TSMC appears to take more capacity on-tap for 10nm and is planning a quick transition from 10nm in 2017 to 7nm in 2018. Samsung, in dissimilarity, is planning to agree at 10nm for multiple product generations. Equally nodes have become more difficult to hitting, each foundry has adopted various methods of defining what constitutes a node shrink. This is nigh apparent with Intel; its 10nm node is expected to exist equivalent to 7nm from TSMC.

Samsung appears to be planning a comparatively late EUV introduction. Its 8LPP will be an 8nm node built with conventional immersion lithography, while its 7nm node introduces EUV in 2019 – 2020, with the CLN7FF+ node.

SamsungFab

Samsung's Austin foundry

Volition EUV be set?

At that place are two things to keep in listen when evaluating these roadmaps. Starting time and foremost, EUV (Farthermost UltraViolet lithography) remains a major confounding variable. TSMC, GlobalFoundries, and Intel all desire to introduce EUV, considering it offers huge potential improvements over double, triple, and quadruple patterning we see today. The problem is (and has been) that EUV is both expensive to deploy and not currently hit the performance targets information technology needs to reach to be a feasible replacement for 193nm lithography.

The second thing to keep in mind is we don't know how well any given company volition execute and we tin can't actually compare nodes based solely on their arbitrary numbering anymore. When TSMC was the only game in town, it was piece of cake to compare their 65nm to 45nm, or 90nm versus 65nm. Now, the state of affairs is much more chaotic, with multiple revisions of each node and each visitor adopting its own naming scheme.

The fact that Intel, TSMC, and GlobalFoundries accept all begun adopting multiple versions of the same node speaks to how difficult information technology is to proceed to deliver consistent performance increases. Ten years agone, the thought of continually iterating on a single procedure node made no sense whatsoever. Today, it's go common identify.

I suspect (though cannot testify) information technology's not just the foundry nodes that get tweaked through this process. At that place's a multi-stride learning procedure in which the foundry tweaks its technology to improve match desired characteristics of a given class of devices, while the developer of said devices (Qualcomm, Intel, AMD, etc) tweaks their design to ameliorate match the process. This was long considered one of Intel'south greatest strengths, though its determination to exit the tablet and smartphone market place more often than not removed it from consideration when comparing with the pure-play foundries.

Again, I recommend giving Anandtech'south total story a look, specially if y'all're curious about promised improvements at various nodes from specific companies.

Now read: The myths of Moore'southward Law