My understanding is that with X-ray etching/lithography it's significantly more difficult to produce the coherent beam and also to do the focusing of it through a masking plate so you can etch a whole wafer at one time. That's the main reason that they've been working on EUV because they can produce a coherent pulse and focus it more easily. For X-rays you have to turn it into a diffraction grating to make the mask and that ends up more difficult to produce (probably not impossible but much more difficult).
Most fundamental technologies take 30 or 40 years to commercialize, if you look at everything from flat screen TVs to digital cameras. The height of the .com bubble was 30 years after the invention of the Internet, and most .com companies didn't actually become profitable until another ten years after that.
I wonder when the rapid erosion of Intel's technical lead in fabrication technology be represented in decreasing evaluation of its market cap.
It was rumored TSMC's 7nm wouldn't be the same as Intel's 10nm. But if they reach ~5nm then they'll likely be knocking on Intel's 10nm door. Combined with the double whammy of 7nm Eypc servers from AMD it seems like Intel's technical offerings are rapidly getting commoditized by the rest of the market.
I always wondered why... I mean I understand the processes will be different for each vendor, but shouldn't 1nm the same for all vendors?
They might use different technologies to achieve it but it should be a measuring unit...
What is the actual difference between Intel and everybody else as to count 10 Intel nm < 7 TSMC nms?
Maybe it's easy to find out in google but cant find the correct combination of keywords...
short answer is there are many parts with sizes in different dimensions x/y/z so what dimension do you measure or average to come up with a final "nm" rating for your cpu?
There realy should be some kind of ISO about node naming.
I guarantee this "5nm" would be as much missleading as their "7nm".
Only thing that matters is wafer price, yield, transistor count and TDP.
So the node scaling was standardized (before ~28nm). It very literally meant MOSfet density (as 28 nm was 1 MOSfet + gap for the next MOSfet). The IEEE had a road map about what to expect out of X sized components, and when those components would be out base on historical data.
But when we hit 20/18/16/14nm that went out the window and it became a marketing term, not so much a literal description. A lot of this was driven by moving to FINfet's which are really MOSfets, as they have lower leakage at smaller sizes, but they also aren't square which makes generalizing a singular node to density a bit wonky.
Well, you can fight the power monster with dark silicon.
Packing things in more tightly lets you spend less time in transit, which might let you squeeze more gates into a cycle, or do the same things a little faster.
Yea for consumers these days, the physical size really doesn't matter much at all. ICs are small enough, nobody's looking forward to 5nm because of the physical size. What matters is cost and electric power usage.
At some level this situation was inevitable, as more and more companies offload their fab capability to TSMC, that gives more and more money for TSMC to invest in boosting their capability, and of course while their margins are a lot less than Intel's with enough money that advantage goes away as well.
Intel's instruction set architecture dominance will keep it going for a long time but ultimately it would probably make sense for Intel to spin off its fabs into the US equivalent of TSMC and capture more margin revenue from their designs versus their process.
50.9% on 32B$ for TSMC[1] vs 62.3% on 62.7B$ for Intel[2]
That is closer than I expected it to be. In terms of actual cash (if you are wondering) Intel ended last year with $22B more of it than TSMC did.
Intel's advantage over AMD has been it's (1) manufacturing technology and (2) architecture, especially since Sandy Bridge.
It turns out that Intel's architecture cheated by skipping privilege checks during speculation, while AMD designs did the correct thing. SMT turns out to be a security nightmare, but in any event AMD now offers the same capability. Either way SMT is another lost advantage.
That leaves Intel's process technology and vertical integration. If it outsources manufacturing it will have effectively ceded these completely, meaning Intel will have lost all of its competitive advantages.
I suppose Intel's human capital might be a competitive advantage, but their missteps cast serious doubt on that.
From what I've heard their isolation has made their tech largely incompatible with external workflows. Similar to how mainframes are technically computers but the software is alien to most people. They don't even use the same names for common things.
That idea of Intel spinning off and developing it's own TSMC sounds great! What other factors have prohibited from someone else developing a more viable TSMC competitor in the US?
The idea isn't exactly unique, in that AMD did it and called it Global Foundries(GF)[1]. Then in 2015 IBM divested their foundry capabilities and sold them to GF as well. It hasn't been a stellar experiment which is largely laid at the feet of the challenge of managing that business.
One of the factors that prevents smaller companies from designing FinFET chips is development cost. Some estimates put the average cost to develop an SoC at around $150 million in labor and IP licenses. With N5 generation, these expenditures will rise to $200 – $250 million, according to EETAsia, which will limit the number of parties interested in using the tech.
Ouch. A $250M NRE cost for an SOC, really puts pressure on guaranteeing the volume of the part in order to recover any sort of margin on a part like that.
Pretty cool to see EUV seem to finally emerge out of the 'someday' status into the 'some now, more next year' status.
Of course, at those prices there will be fewer SOCs. That means, on average, that each SOC will get more design wins, and therefore have higher volume...
It also means that each SoC will have more pressure to cater to more customers which seems more dark silicon which will reduce the advantage of smaller process sizes
What is "risk production"? I can't help but reading it as something like, "we even risk producing some chips at some point!", which I presume isn't quite it.
>> "Risk Production means that a particular silicon wafer fabrication process has established baseline in terms of process recipes, device models, and design kits, and has passed standard wafer level reliability tests."
It sounds like a batch produced to certify the process for insurance and contract purposes.
That's sort of right, basically it's the first real production runs, when you think everything is ready & will work, but you haven't actually made a real product yet.
At the start of risk production, by definition there have been no customer designs put through the fab. Actually this is not quite true, Sun explained. Before the start of risk production the company has already run a number of shuttles with test chips from customers, so foundry and clients are already starting to wring out the more critical structures in the first customer designs. But these test shuttles are not full chips either.
As a customer I've always heard it used as the latter when we tape out (or spin from a bug) and think we're ready to go so we start some wafers down the line behind the first ones to reduce time to market
From starting a batch to where products are actually testable takes >3 months. "Risk production" means that the foundry says, "okay, we think everything is fine now, but we make no guarantees that it will work". Customers then have the option of purchasing super expensive wafer starts that might or might not work. You might luck out and have lots of great next-gen chips months before your competitors, or alternatively you might get the GF100 thing happen to you where you get five working chips out of each $10k wafer.
> By contrast, TSMC’s second-generation 7 nm manufacturing technology (CLN7FF+, N7+) will use extreme ultraviolet lithography for four non-critical layers
As in, "like everyone else, we haven't really figured out EUV shot noise yet".
Well the solution is simple: use longer exposures, of course, that has the unpleasant side effect of annihilating your throughput on extremely expensive machines...
Just to help me clarify, the tariffs that are being levied between China and America, do or do not also apply to Taiwan? I’ve never fully understood how the separation of Taiwan and China is made since it seems to be such a touchy subject.
The RoC consider themselves the rightful "China" too. Some Taiwanese politicians have advocated declaring themselves an independent country but that's never been the official government position.
In general, the US maintains unofficial diplomatic relations with Taiwan, which means that although the US does not officially recognize the government of Taiwan, it is subject to different tariffs than those imposed on China. Similarly, tariffs imposed by Taiwan are set by the ROC, not the PRC, and therefore are totally distinct from this trade war.
More information can be found in this position statement from the Department of State:
Interesting, so in that case, does Taiwan benefit from this ongoing trade war, or is the heightened tension and fear of political/military escalation nullifying any benefit that might be felt?
Taiwan is hurt indirectly. Taiwan has invested heavily on the mainland China and manufactures there.
Many Chinese products exported to US have very low Chinese value added. Chinese manufacturer buys Taiwanese, Japanese or US made components, assembles them and sell them to the US.
For many high-tech products, Chinese value added may be just 20% of the price of the product. Chinese tariffs hit everyone in the value chain. Even the US.
Just after World War II, the Communists took over China from the previous government. The previous government ran to the island of Taiwan, which had been part of China. If I understand correctly, you now have two different governments, one on the mainland and one on Taiwan, both of which claim to be the government of China, and both of which claim to rule all of China, and both of which claim that there is only one China.
Other nations play pretend to various degrees, trying to placate (mainland) China without too much swallowing their own pride in not being China's lapdogs. But they're effectively two separate countries.
So: The tariffs being applied to China do not apply to Taiwan. However, Trump seems to be throwing tariffs at several targets lately; he may or may not have applied some to Taiwan.
The current government of Taiwan does not claim sovereignty over the mainland. They still call themselves "Republic of China" though. Indeed most countries do not recognize Taiwan to placate the mainland regime which adheres to the "One China" policy of threatening war over official separation by Taiwan.
Sometimes you wonder how things can stay broken for so long.
> If Hong Kong had been militarily defensible, it'd be in the same situation.
If the UK had really wanted to keep Hong Kong, could it have militarily defended it? No doubt the geography is less than favourable; but, keeping in mind that UK and PRC are nuclear powers, would regaining Hong Kong be worth the risk of war between two nuclear states? PRC might well have decided it wasn't worth that risk.
But in reality, the UK had no real interest in keeping Hong Kong, and were more than happy to hand it over to PRC. (Yes, there was a treaty saying they had to give back a big chunk after 99 years, but if the UK really wanted to get out of that, they could have found a solution – e.g. if they hadn't switched recognition from ROC to PRC, they could have asked ROC for a lease extension. Or, they could have made their switch of recognition from ROC to PRC conditional on PRC granting them perpetual sovereignty over the whole of Hong Kong.)
> If the UK had really wanted to keep Hong Kong, could it have militarily defended it?
The Chinese would not have to have fired a shot to make HK surrender. HK was then a city with more than 6 million citizens, which was entirely dependent on water imported from the mainland. The treaty that allowed the importation of all that water was only set until the end of the lease, and the Chinese very clearly stated that they had no interest in extending it.
Also, earlier when HK expressed some interest in investing into desalinization, the Chinese threatened to immediately cut off the water supply if it seemed like the HK local government would have actually tried to build up enough water resources to make them not dependent on the mainland.
The result of shutting down the taps would have been millions dead within a week. Keeping HK British would have required a massive invasion of China by the UK, just to secure water. I don't see that as very likely.
China did not control the island before the refugees from China took it over. Japan controlled the island. However, after the war Japan did hand control over the island to China however China never did get to control it after the war or before.
Its up to the US who to put its own import tariffs on, so the US can just say "imports from China" but not directly from Taiwan.
China may be able to force import tariffs on goods imported by Taiwan (idk not an expert) but I bet Taiwan imports relatively little directly from the US?
Samsung is going mass production with their EUV in Q1 or 2 in 2019. I'm also speculating that Qualcomm's next SD would be made by Samsung LSI. How does it compare to TSMC?
Does someone have an up-to-date table comparing the different fabs' interpretations of "Xnm" process size? I recall them all having very different values (to the point where one might actually be a whole process size behind, but not admit it for marketing purposes).
I remember when I was doing my master's thesis for ASML in Eindhoven in Holland (their headquarters). They had a separate building on the campus just for the development of EUV, which was the next new thing. This was in 2011 and they had already been working on it for a couple of years. Happy to see that chip manufacturers are now using this technology for their production. It looks like the bet paid off!
61 comments
[ 3.4 ms ] story [ 91.7 ms ] threadIt was rumored TSMC's 7nm wouldn't be the same as Intel's 10nm. But if they reach ~5nm then they'll likely be knocking on Intel's 10nm door. Combined with the double whammy of 7nm Eypc servers from AMD it seems like Intel's technical offerings are rapidly getting commoditized by the rest of the market.
Maybe it's easy to find out in google but cant find the correct combination of keywords...
https://news.ycombinator.com/item?id=17854700
But when we hit 20/18/16/14nm that went out the window and it became a marketing term, not so much a literal description. A lot of this was driven by moving to FINfet's which are really MOSfets, as they have lower leakage at smaller sizes, but they also aren't square which makes generalizing a singular node to density a bit wonky.
Wouldn't 5 nm allow shorter wire traces, which would reduce another limiting factor in computational speed?
The fundamental limit of clock speed is power draw. As clocks increase the wattage ~ frequency relation goes from
It starts becoming As you start getting >3GHz so while we can make processors that run >5GHz. There just aren't applications that benefit a lot from it.Packing things in more tightly lets you spend less time in transit, which might let you squeeze more gates into a cycle, or do the same things a little faster.
Intel's instruction set architecture dominance will keep it going for a long time but ultimately it would probably make sense for Intel to spin off its fabs into the US equivalent of TSMC and capture more margin revenue from their designs versus their process.
[1] http://www.tsmc.com/download/ir/annualReports/2017/english/p...
[2] https://s21.q4cdn.com/600692695/files/doc_financials/2017/an...
It turns out that Intel's architecture cheated by skipping privilege checks during speculation, while AMD designs did the correct thing. SMT turns out to be a security nightmare, but in any event AMD now offers the same capability. Either way SMT is another lost advantage.
That leaves Intel's process technology and vertical integration. If it outsources manufacturing it will have effectively ceded these completely, meaning Intel will have lost all of its competitive advantages.
I suppose Intel's human capital might be a competitive advantage, but their missteps cast serious doubt on that.
[1] https://www.globalfoundries.com/
One of the factors that prevents smaller companies from designing FinFET chips is development cost. Some estimates put the average cost to develop an SoC at around $150 million in labor and IP licenses. With N5 generation, these expenditures will rise to $200 – $250 million, according to EETAsia, which will limit the number of parties interested in using the tech.
Ouch. A $250M NRE cost for an SOC, really puts pressure on guaranteeing the volume of the part in order to recover any sort of margin on a part like that.
Pretty cool to see EUV seem to finally emerge out of the 'someday' status into the 'some now, more next year' status.
>> "Risk Production means that a particular silicon wafer fabrication process has established baseline in terms of process recipes, device models, and design kits, and has passed standard wafer level reliability tests."
It sounds like a batch produced to certify the process for insurance and contract purposes.
At the start of risk production, by definition there have been no customer designs put through the fab. Actually this is not quite true, Sun explained. Before the start of risk production the company has already run a number of shuttles with test chips from customers, so foundry and clients are already starting to wring out the more critical structures in the first customer designs. But these test shuttles are not full chips either.
-- https://www.edn.com/electronics-blogs/practical-chip-design/...
So, it's a point in the process where the foundry is taking a chance, and the customer is also taking a chance.
I've also heard it applied to a customer design, e.g. the customer starts to volume ramp production before they finish fully testing the design.
As in, "like everyone else, we haven't really figured out EUV shot noise yet".
Nice marketing though, lolz.
Octa patterning? ;)
The wiki page on EUV is uncharacteristically excellent.
https://en.wikipedia.org/wiki/Extreme_ultraviolet_lithograph...
The mainland Chinese government disagrees.
I don't know about these particular tariffs, but I believe the U.S. would not consider tariffs aimed at (mainland) China to apply to Taiwan.
There are things that are de facto true, even if they're not de jure true.
More information can be found in this position statement from the Department of State:
https://www.state.gov/r/pa/ei/bgn/35855.htm
Taiwan is hurt indirectly. Taiwan has invested heavily on the mainland China and manufactures there.
Many Chinese products exported to US have very low Chinese value added. Chinese manufacturer buys Taiwanese, Japanese or US made components, assembles them and sell them to the US.
For many high-tech products, Chinese value added may be just 20% of the price of the product. Chinese tariffs hit everyone in the value chain. Even the US.
Other nations play pretend to various degrees, trying to placate (mainland) China without too much swallowing their own pride in not being China's lapdogs. But they're effectively two separate countries.
So: The tariffs being applied to China do not apply to Taiwan. However, Trump seems to be throwing tariffs at several targets lately; he may or may not have applied some to Taiwan.
Sometimes you wonder how things can stay broken for so long.
Taiwan doesn't want the mainland's government, and the mainland doesn't want Taiwan's government.
If Hong Kong had been militarily defensible, it'd be in the same situation.
If the UK had really wanted to keep Hong Kong, could it have militarily defended it? No doubt the geography is less than favourable; but, keeping in mind that UK and PRC are nuclear powers, would regaining Hong Kong be worth the risk of war between two nuclear states? PRC might well have decided it wasn't worth that risk.
But in reality, the UK had no real interest in keeping Hong Kong, and were more than happy to hand it over to PRC. (Yes, there was a treaty saying they had to give back a big chunk after 99 years, but if the UK really wanted to get out of that, they could have found a solution – e.g. if they hadn't switched recognition from ROC to PRC, they could have asked ROC for a lease extension. Or, they could have made their switch of recognition from ROC to PRC conditional on PRC granting them perpetual sovereignty over the whole of Hong Kong.)
The Chinese would not have to have fired a shot to make HK surrender. HK was then a city with more than 6 million citizens, which was entirely dependent on water imported from the mainland. The treaty that allowed the importation of all that water was only set until the end of the lease, and the Chinese very clearly stated that they had no interest in extending it.
Also, earlier when HK expressed some interest in investing into desalinization, the Chinese threatened to immediately cut off the water supply if it seemed like the HK local government would have actually tried to build up enough water resources to make them not dependent on the mainland.
The result of shutting down the taps would have been millions dead within a week. Keeping HK British would have required a massive invasion of China by the UK, just to secure water. I don't see that as very likely.
China may be able to force import tariffs on goods imported by Taiwan (idk not an expert) but I bet Taiwan imports relatively little directly from the US?
https://en.m.wikipedia.org/wiki/7_nanometer
https://en.m.wikipedia.org/wiki/10_nanometer