E-Cat
Member
UPDATE 1:
---
DISCLAIMER: This analysis is mostly based off of my own intuition from reading various investor briefings. Also, I realize that TSMC is not the only foundry and that AMD have thus far used GloFo's 14nm due to their contractual obligations. However, I have included some ASML remarks, which it should be clear that they have wider reaching implications for the whole industry. After all, they make the machines that the foundries use to make their machines. Anyway, I actually hope to be proven wrong on this theory and will continue to update the OP as relevant information becomes available.
There is a precedent from recent history, i.e., the 20/16nm situation, where we were stuck on 28nm for 4 years and 4 months before NVIDIA launched the GTX 1080 in May. 20nm was the last planar node before FinFET, and for good reason, as there was a lack of power scaling to match the increased transistor density; thus, raising TDP and voltage issues that made 20nm GPUs a non-starter.
Conversely, the issue with 10nm may not have to do with the underlying technology: http://www.tsmc.com/uploadfile/ir/quarterly/2015/3C2bO/E/TSMC 3Q15 transcript.pdf
Indeed, the node appears very much in line with your typical power and density requirements for a Moore's law like progression.
So why bring up 20/16nm? Well, it turns out there are some similarities. Just as the 16nm manufacturing technologies utilized the same planar back-end interconnect originally designed for 20nm, TSMC are planning to leverage the 10nm tool compatibility and maturity from volume production with the 7nm process; essentially treating the 10/7nm dependence analogously to 20/16nm: http://www.tsmc.com/uploadfile/ir/quarterly/2016/1yk4i/E/TSMC 1Q16 transcript.pdf
Being a small node means a limited number of wafers: http://seekingalpha.com/article/398...-results-earnings-call-transcript?part=single
Having a limited wafer capacity means that the big customers at the top of the pecking order will gobble it all up. This means Apple, et al. They set up the cadence where new A1x chips are needed in volume for the new iPhone launch, which tends to be September. So that means volume ramp in the first half of the year, which means tapeout/risk production the year prior. The GPU companies get the scraps typically one year after the mobile manufacturers.
But now it's one year after and everybody's moving to 7nm! The production yield ramps are fantastic on the tried and tested tools, and wafer prices are plummeting. Meanwhile, the remaining 10nm capacity has dried up and is only used for some esoteric applications where the margin is irrelevant.
TSMC in their quarterly calls went from treating 10/7nm as a monolithic lump to relegating 10nm as a primarily 'mobile node'. Very telling: http://www.tsmc.com/uploadfile/ir/quarterly/2015/4T5Um/E/TSMC 4Q15 transcript.pdf
that basically all the capacity will be spent fulfilling the mobile manufacturers' orders;
then, everybody rapidly moves on to 7nm which is tracking unusually close to 10nm with
excellent yield/power/density characteristics and is being aggressively adopted by
high-performance computing, mobile GPU and game console people.
The 10nm node is expected to be a short lived one, created in large part for Apple’s iPhone 8. It sports twice the gate density, and either 10 percent higher speed or 25 percent less power than the 16nm node, TSMC said.
---
DISCLAIMER: This analysis is mostly based off of my own intuition from reading various investor briefings. Also, I realize that TSMC is not the only foundry and that AMD have thus far used GloFo's 14nm due to their contractual obligations. However, I have included some ASML remarks, which it should be clear that they have wider reaching implications for the whole industry. After all, they make the machines that the foundries use to make their machines. Anyway, I actually hope to be proven wrong on this theory and will continue to update the OP as relevant information becomes available.
There is a precedent from recent history, i.e., the 20/16nm situation, where we were stuck on 28nm for 4 years and 4 months before NVIDIA launched the GTX 1080 in May. 20nm was the last planar node before FinFET, and for good reason, as there was a lack of power scaling to match the increased transistor density; thus, raising TDP and voltage issues that made 20nm GPUs a non-starter.
Conversely, the issue with 10nm may not have to do with the underlying technology: http://www.tsmc.com/uploadfile/ir/quarterly/2015/3C2bO/E/TSMC 3Q15 transcript.pdf
TSMC said:"This technology [10nm] has a logic density of 2.1 times of its previous generation, that's 16 FinFET Plus, with performance of 20% enhancement or a power consumption reduction of 40%."
Indeed, the node appears very much in line with your typical power and density requirements for a Moore's law like progression.
So why bring up 20/16nm? Well, it turns out there are some similarities. Just as the 16nm manufacturing technologies utilized the same planar back-end interconnect originally designed for 20nm, TSMC are planning to leverage the 10nm tool compatibility and maturity from volume production with the 7nm process; essentially treating the 10/7nm dependence analogously to 20/16nm: http://www.tsmc.com/uploadfile/ir/quarterly/2016/1yk4i/E/TSMC 1Q16 transcript.pdf
Why is this important? One reason is that due to 95% of the equipment being used for 10nm being compatible with 7nm, the planned risk production for 7nm is able to commence only five quarters after 10nm--the typical time for such a transition being closer to two years. In other words, 10nm is going to be a very short-lived node in the echelon of nodes.TSMC said:"Our 7-nanometer technology... is a further extension of N10 technology, with more than 60% in logic density gain and 30% to 40% reduction in power consumption. N7 fully leverages N10 learning and shares more than 95% of common tools."
Being a small node means a limited number of wafers: http://seekingalpha.com/article/398...-results-earnings-call-transcript?part=single
ASML said:"But what we’re currently planning is that, like you pointed out, the 10 nanometer node is not a very large node. It will not entirely go away. The expectations for the 7 nanometer node are actually quite strong. So there is always going to be some level of reusage actually planned. We don’t think that percentage reuse is going to be anything significantly different than what we saw in the past".
"So it really depends on whether the installed capacity for the 10 nanometer node currently has enough customer base. And you have to remember, it is not a lot. The 10 nanometer installed by the end of this year, maximum 35,000, 40,000k, so that’s not a lot. So you have some big customers and you’re full and then everybody moves to 7, which will be different customers. So it’s difficult to say very customer specific, very specific to their customer base. And what we’re currently planning and also in our longer-term planning is that we don’t see as a major shift or a major impact of that reuse model going forward."
Having a limited wafer capacity means that the big customers at the top of the pecking order will gobble it all up. This means Apple, et al. They set up the cadence where new A1x chips are needed in volume for the new iPhone launch, which tends to be September. So that means volume ramp in the first half of the year, which means tapeout/risk production the year prior. The GPU companies get the scraps typically one year after the mobile manufacturers.
But now it's one year after and everybody's moving to 7nm! The production yield ramps are fantastic on the tried and tested tools, and wafer prices are plummeting. Meanwhile, the remaining 10nm capacity has dried up and is only used for some esoteric applications where the margin is irrelevant.
Knowing all this; if you're NVIDIA or AMD, do you even design an architecture for 10nm or skip straight to 7nm? We know that Volta is almost certainly going to be 16nm; Navi is a bigger question mark as it's coming in 2018 when you would expect to see 10nm, if ever. More likely it will be 7nm.ASML said:"As we know, the 10 nanometer ramp for foundry is, in fact, a part of a, let’s say, bigger logic node, what you call, the 10/7 nanometer node. So what we will see in 2017, the 7 nanometer node is expected to be as a strong and a large node. We expect that, in 2017, the continuation as almost a logical evolution from 10 nanometer into the 7 nanometer node to happen."
TSMC in their quarterly calls went from treating 10/7nm as a monolithic lump to relegating 10nm as a primarily 'mobile node'. Very telling: http://www.tsmc.com/uploadfile/ir/quarterly/2015/4T5Um/E/TSMC 4Q15 transcript.pdf
<-- Notice the mention of 'game console'? Hmm, interesting. However, one quarter later: http://www.tsmc.com/uploadfile/ir/quarterly/2016/1yk4i/E/TSMC 1Q16 transcript.pdfTSMC said:"These two technologies, 10-nanometer and 7-nanometer, will cover a very wide range of applications, including application processors for smartphone, high-end networking, advanced graphics, field-programmable gate array, game console, wearables and other consumer products".
'Game console' now appears solely on 7nm. No mention of 'advanced graphics' or 'high-performance computing' on 10nm. By Q2, everything points to a graceful mass exodus: http://www.tsmc.com/uploadfile/ir/quarterly/2016/2izbs/E/TSMC 2Q16 transcript.pdfTSMC said:"Let me give you first on N10 update. We have received N10 customer product tape-out in 1Q 2016. We are actively preparing for more customer product tape-outs in the following quarters. Most of our N10 [=10nm] users are for mobile products. We will put this technology in production in two of TSMC's 12-inch giga-fabs. Those tape-outs will drive a sizable demand starting from 2Q 2017 through 2018."
"We have expanded our N7 design ecosystem development to include both mobile and high-performance computing, to enable our customers to deliver their first-to-market products. Our N7 adoption is very strong, with customers ranging from mobile GPU, game console, FPGA, network processors and other consumer product applications. We have more than 20 customers in intensive design engagement with us and expect to have 15 customer tape-outs in 2017. The volume production of N7 will start from first half 2018."
tl;dr: 10nm is bound to be a small node as revealed by credible entities - reasonable to inferTSMC said:"Our 10-nanometer has been transferred from R&D to production. Our first 10-nanometer customer product has been produced with satisfactory functional yield. So far, three customer products have been taped out to us. More customer product tape-outs are expected later this year. Those product tape-outs will start a revenue stream starting first quarter 2017, which will ramp steeply throughout 2017."
"Our 7-nanometer technology development is well on track. Its 256-megabit SRAM yield improvement is ahead of our schedule. In addition, we believe our 7-nanometer PPA, that is power, performance and area density, with its schedule, is ahead of our competitors. This technology has been aggressively adopted, not only by mobile customers, but also by high-performance computing customers. They all have aggressive product tape-out plan in first half 2017, with volume production planned in early 2018."
that basically all the capacity will be spent fulfilling the mobile manufacturers' orders;
then, everybody rapidly moves on to 7nm which is tracking unusually close to 10nm with
excellent yield/power/density characteristics and is being aggressively adopted by
high-performance computing, mobile GPU and game console people.