Every Government wants a fab, Does the market?

[Arthur Hanson]

Governments local and national all want fabs and most have little to no knowledge of the industry, especially on a worldwide scale. This can and will lead to some ugly distortions in the market and its supply chain. There will definitely be some wrong mixes both in supply and demand that will have both interesting and terrifying consequences economically and politically. Companies and facilities that are late to the game and not well thought out will get slaughtered. Politicians should be doing due diligence, but most aren't. Will educational institutions in these areas make the same mistakes? Any thoughts and comments on this appreciated.

 

[hist78]

"Those governments everywhere can promise you the subsidies but cannot guarantee you the market demands."

Post in thread 'UMC Reports Sales for June 2023' https://semiwiki.com/forum/index.php?threads/umc-reports-sales-for-june-2023.18335/post-62255

 

[Tanj]

In the short term the fabs will need to hustle to have what customers need. Not just chips, but also shorter cycle times, open PDKs, qualify 3rd party IP libraries, qualify and package for special markets, run smaller batches efficiently, support startup designs. Rethink how to make an ordinary sized customer welcome, supported, and cost effective. Things an oligopoly begrudges.

In the long run the market may grow faster with more capacity and a better customer orientation.

 

[Daniel Nenni]

Long term governments can turn empty fab shells into affordable housing.

It will be interesting to see how quickly fabs can be built in different countries. Will it take longer to build a fab in the US? And how much more will it cost? We could see some fab shaming coming our way.

 

[IanD]

There will undoubtedly be problems as governments around the world try to bring up local fabs, some of which Arthur mentioned.

However these are probably small beer compared to the problems if everything advanced is left in Taiwan and China invades -- sorry, has a "special re-unification military operation" -- and then China holds the worldwide electronics industry -- sorry, industry in general -- to ransom, or shuts down parts of it they don't like.

And before anyone says "China won't do that, it would be stupid beyond belief, it would be disastrous for them" -- well, that's what they said about Russia before it invaded Ukraine.

The chances of this happening might not be huge (but seem to be increasing), but the consequences would be -- so it makes sense to take out insurance against it, and that's what local fabs are. Even if the insurance premiums are high, governments are now deciding it's worth it for the sake of national security.

 

[hist78]

There will undoubtedly be problems as governments around the world try to bring up local fabs, some of which Arthur mentioned.

However these are probably small beer compared to the problems if everything advanced is left in Taiwan and China invades -- sorry, has a "special re-unification military operation" -- and then China holds the worldwide electronics industry -- sorry, industry in general -- to ransom, or shuts down parts of it they don't like.

And before anyone says "China won't do that, it would be stupid beyond belief, it would be disastrous for them" -- well, that's what they said about Russia before it invaded Ukraine.

The chances of this happening might not be huge (but seem to be increasing), but the consequences would be -- so it makes sense to take out insurance against it, and that's what local fabs are. Even if the insurance premiums are high, governments are now deciding it's worth it for the sake of national security.

The supply chain diversification is a great idea but there are opportunists and too much wishful thinking. I'm not sure if there are minimum capacity, minimum output volume, and profitability requirement in the government subsidies agreement. If there is no party pooper to raise these types of hard questions today, we will end up to see those new fabs keep asking for additional public money again and again for the critical goals they have no ability to fulfil.

This can severely hurt the trust and interest that people and government may have on the semiconductor industry in the long run.

 

[Barnsley]

There will undoubtedly be problems as governments around the world try to bring up local fabs, some of which Arthur mentioned.

However these are probably small beer compared to the problems if everything advanced is left in Taiwan and China invades -- sorry, has a "special re-unification military operation" -- and then China holds the worldwide electronics industry -- sorry, industry in general -- to ransom, or shuts down parts of it they don't like.

And before anyone says "China won't do that, it would be stupid beyond belief, it would be disastrous for them" -- well, that's what they said about Russia before it invaded Ukraine.

The chances of this happening might not be huge (but seem to be increasing), but the consequences would be -- so it makes sense to take out insurance against it, and that's what local fabs are. Even if the insurance premiums are high, governments are now deciding it's worth it for the sake of national security.

It depends , Singapore for instance , a lot of the companies here are either building new FAB , expanding or re-utilising buildings.

None of these are Singaporean comapanies and I would doubt this is being done for the sake of National Security.

 

[Barnsley]

The supply chain diversification is a great idea but there are opportunists and too much wishful thinking. I'm not sure if there are minimum capacity, minimum output volume, and profitability requirement in the government subsidies agreement. If there is no party pooper to raise these types of hard questions today, we will end up to see those new fabs keep asking for additional public money again and again for the critical goals they have no ability to fulfil.

This can severely hurt the trust and interest that people and government may have on the semiconductor industry in the long run.

One would assume that those companies expanding capacity have some writtem gaurantees in place for at least a couple of years of business from their customers, otherwise its quite the gamble.

 

[Mickey]

"Those governments everywhere can promise you the subsidies but cannot guarantee you the market demands."

Post in thread 'UMC Reports Sales for June 2023' https://semiwiki.com/forum/index.php?threads/umc-reports-sales-for-june-2023.18335/post-62255

Seems you are quite right.

One would assume that those companies expanding capacity have some writtem gaurantees in place for at least a couple of years of business from their customers, otherwise its quite the gamble.

Politicians and global leaders all want fabs to mitigate and steer clear of geopolitical tensions. The harsh decisions were not well taught out by the lessons learned from the turbulent and bloody history of the flourishing semiconductor fabrication industry just a few decades ago.

It appears most of such de-risk moves may invoke much higher risk of market or demand failures then ever before. Let’s wait and see.

 

[delong.height]

Government policies will be the most important driving force for these fabs' success. When these fabs are build, I guess we will see a shift in government policy to be more protectionism in order to make sure the companies will continue to invest and expand the fabs, such as Intel, Samsung, Micron, etc. Currently it's not visible as these fabs are still in construction and to not create more headaches or pain for the industry. Chips Act can be considered as one, but they are not as effective and organic as fundamental as tariffs and industrial policies.

But when it's at right time, governments around the world will mandate that these fabs need to be 90% filled, a certain percentage of advanced semiconductor to be made in their respective countries, a higher tax rate to be imposed on companies that has massive supply chain outside their country (such as Apple), an annual commitment for expanded capacity, etc. And just imagine what could happen in next election, we may get Trump again. So, government policies will be the more important and easier way to grow the semiconductor foundry industry.

People should not be afraid of capacity being wasted. In the end, the demand for better compute always exceeds supply in just a few years. Not only that, but it will also create massive opportunities for entire supply chains under these foundries, like TSMC, Intel, and Samsung as these companies have to spend more to build more capacity that may seem unnecessary at the time, and fabless companies that can have more options to choose from.

 

[Xebec]

(This is oversimplified, but.)

I think at some point governments will just realize it may be cheaper and easier to stockpile advanced FPGAs and a lot of general semiconductor supplies. Like ammunition..

So if a serious disruption occurs, all they need to do is fabricate PCBs (and program the FPGAs) and they at least will have some ability to keep critical systems up and running, giving them time to find trade with someone else who can make chips.

 

[Tanj]

I like FPGAs but they are FAR from replacing ASICs in general. The only reason they compete in things like inferencing is the heavy investment in hard IP for the specific market., where they add effectively adaptive smart routing to the compute blocks. So they might, in a pinch, do something useful if you need to replace an inferencing ASIC.

For other kinds of ASIC there is not that lucky dominant set of hard IP blocks over a wide enough market segment for there to be FPGAs tuned well enough to be replacements for a good ASIC. Maybe a new FPGA for a very old ASIC. Maybe.

In general there needs to be innovation in the agility of Fabs and the ability for medium or small Fabs to be commercially interesting. So you might be able to make, say, a 7nm "rapid replacement" version of a wide range of ASICs rather better than FPGAs. Maskless lines and fab processes tuned for short overall cycle times. A lot of EDA and compatible IP librariies needed for a variety of analog functions, too.

Really, a huge step from where we are now, or even where this first round of geographic diversity will get us.

 

[nghanayem]

In general there needs to be innovation in the agility of Fabs and the ability for medium or small Fabs to be commercially interesting. So you might be able to make, say, a 7nm "rapid replacement" version of a wide range of ASICs rather better than FPGAs. Maskless lines and fab processes tuned for short overall cycle times. A lot of EDA and compatible IP librariies needed for a variety of analog functions, too.

Really, a huge step from where we are now, or even where this first round of geographic diversity will get us.

I don't think this is possible (unless you want to pay a dozen orders of magnitude more per wafer). I suspect not having high tool utilization, economics of scale, or using non HVM E-beam litho tools instead of DUV/EUV tools would lead to wafer costs so astronomically high you would be better off buying heaps of general purpose xPU to get the necessary compute. Obviously I couldn't even begin to give you an actual estimate for the cost adder of doing this, but the one thing I can say with certainty is that this idea would go against engineering and semiconductor fundamentals. I would even wager that ASICs on super easy to design for node like 90nm would be more compelling for this use case than a hypothetical Cliff-MC "7nm" ASIC node (even if it pains me to say some lame 90nm thing would be better than some cool E-beam litho node running in a fab where lots never wait for loadports).

 

[Tanj]

I believe ebeam is feasible at a few wafers per hour. Maskless, and with good margins. I gave a talk on my approach at Photomask Japan 2023. However, that does not get you the whole chip, because the resist chemistry is same as EUV and no process is designed to use that above roughly M3, which means that there is a question how to get maskless to be cost effective with DUV at a few wafers per hour so that the whole stack is approached in a uniform way. A mismatched production line is not going to perform well.

Indeed, how the other parts of the process adapt to short total cycles and small to medium batches is also interesting. I have not found a lot of clarity in how that works, although clearly some machines have significant setup times for different batches, and other parts of the process have feedback loops incorporating metrology.

The publicity around the formation of Rapidus indicates they are interested in small batch and short cycle but it is hard to know exactly what they are thinking. I know Japan has funded research on small lines but what was published has not seemed useful outside of a research lab.

 

[cliff]

I'm with you Tanj. If I make a sea of logic gates and bring them up to met2 in 180nm, can you sell me one of dem embeamers for our office? One 6" wafer per day per day would be great. Do you have a copper 3D printer? That would work too.

 

[dlip410]

I may have missed it, but no one has considered the inevitable dilution of tooling and materials that would ensue. It seems to be a difficult situation now! What will happen if there are 2-3x Fabs to support in the near future?

 

[cliff]

Dlip, I don't agree with you. I think the problem is that we don't have enough 65'ish nm fabs with lots of metal layers and quick turnaround times as Tanj has suggested. In the 90's, many companies had their own fabs. The industry (like all industries) have consolidated into oligopolies. Fast turnaround foundries/methods at 65-40nm at small volume with breakout boards that allow tight I/O pads to expand to something easier to interpose/hybridize would be huge, rather than relying on these miserably inefficient FPGAs.

Note: I am making an embedded 4 letter word because of their industry acceptance and the lack of fast turnaround time fabs. I really don't want to, but the slow cycle times and huge mask costs require it. Tanj is on the money. I hope he succeeds, but I gotta believe what he is doing is difficult. My stuff far less ambitious. It is virtual. If you can't beat them...

We need a fast turnaround cheapskate fab. I tried calling LA Semiconductor (former ONSemi 180 or so), but they don't answer. Do you really think they will dilute the industry?

 

[dlip410]

Hi Cliff,
Yes, I agree that legacy ME should be a priority. For instance, I think that the world is incredibly vulnerable if China decides to disrupt the supply of 28nm ME. A global agenda that addresses the production of legacy ME is just as important as SOTA ME. Unfortunately, it’s not a sexy endeavor, doesn’t drive job creation, and wont get anyone re-elected.

 

[Tanj]

.. the problem is that we don't have enough 65'ish nm fabs with lots of metal layers and quick turnaround times as Tanj has suggested.

While I appreciate the shout out, I am not a fan of 65nm as a general case. Like most of the surviving older processes it has evolved to offer special advantages which may work for a given project, but in general it is a niche. If you have a customer base for quick-turn chips where 65nm works fine, then great! Relatively low cycle time and cheap. I know some chips used in servers stalled at 65nm for exactly those reasons. They could spin you a custom variant ASIC in less than 6 months. Could be a lot of embedded or other markets do the same, stall anywhere down to 28nm or 16ff then iterate with familiar tools while the process node is good enough.

My comments that may have been read that way have to do with putting the upper metals on with DUV, but the modern processes at 7nm and finer in the low levels are going to have the most valuable uses, for good reason - energy per operation.

A 5nm FinFET operates at less than half the voltage of a 65nm device and is loaded with a small fraction of the capacitance. Modern applications consume TOPs at fractions of a pJ per operation, doing everything from digital radio to sensory fusion to speech reco to putting cat ears on your video chat or Tik Tok to self driving .. etc. A world 65nm will never serve the trillions of ops per second that modern apps take for granted.

 

[nghanayem]

I believe ebeam is feasible at a few wafers per hour.

That's the problem. Even an array of 10 of them would have comparable throughputs to prototype EUV tools from the 2000s. Even assuming 5 wph, the E-beam tools would take up ~30x more space than a comparable DUV tool and ~5x the size of the new hotness of low NA EUV tools. A 20k WSPM shell would now have like 2000 WSPM even if I am extremely generous with the napkin math and assume the non litho tools take up negligible floor space (considering the swarm of E-beam tools needed that might not even be that bad of an assumption). I'd be more than willing to wager that 10 Genoas or 10 Nvidia hoppers would blow away one of your ASICs made on this 7"nm" process and not bankrupt the foundry while doing it.

Maskless, and with good margins. I gave a talk on my approach at Photomask Japan 2023. However, that does not get you the whole chip, because the resist chemistry is same as EUV and no process is designed to use that above roughly M3, which means that there is a question how to get maskless to be cost effective with DUV at a few wafers per hour so that the whole stack is approached in a uniform way. A mismatched production line is not going to perform well.

Using DUV would also defeat the stated purpose of it being maskless flow so anybody in their garage can tape out on this thing.