Really enjoyed reading this article, absolutely awesome job. As a retired EE with some experience with IC design and fab (back in the 70s), I can really appreciate the engineering innovation that goes into a modern fab.
I can’t help but wonder how the engineering challenge of a modern fab compares to, say, an AP1000 nuclear plant. Seems like a nuclear plant is much, much simpler. Nuclear costs are probably around 5 times what they should be due to over wrought safety concerns.
A new fab to make 5nm devices or smaller on 300mm wafers if a near miracle. But if you want to know a system much more complex, look at the human body, biochemistry, aging, the biosphere that are much more complicated and much harder to understand comprehensively. However, faster chips, more data storage, faster telecom, better SW, AI, perhaps quantum computing at some point, could all contribute to making a greens, safer, healthier planet for 8 billion - profitably. Let's go beyond More Moore, with Semicon and Technology, photonics, EUV etc help.
Despite these incredible complications and requirements, my Apple Studio is no faster using CPU/Graphics intensive programs like Sketch-Up than my 2015-manufactured iMac I replaced it with. The new machine uses a Apple M2 Ultra with 64GB RAM, double the RAM of the old machine, but without the Intel CPU and Nvidea GPU. Maybe these old chipmakers knew something that Apple does not when it replaced their chips with a combo chip of its own, or maybe Apple sought to cut costs and sacrificed function. The new machine cost about what the old one did, despite 8 years of inflation between them, with extra cost because the monitor is not separate if you want 27", which I did (the newer iMacs only come in 24" models).
My major model is a 380MB Sketch-Up model building, but that was really close to 800MB before Sketch-Up compressed the filesize in their latest desktop version (and last; now they want a yearly subscription for the newer versions). It is noticeably slower to load, manipulate, etc., especially if other things are running. If I had known that up front, I would have just replaced the hard drive in the old machine when that died. In contrast, the 2015 iMac was much faster than the 2009 iMac it replaced, which was so slow it became unusable for my model and had to be replaced.
Apple, meanwhile, just announced a $100b stock buyback, on top of similar buybacks last year. It's clear they are not investing the same money into computers as previously, and even the endless iPhone upgrades are just evolutionary, not revolutionary.
Nothing can accelerate an industry that has decided to stagnate.
A video fly-through is here: https://bit.ly/Riverarch, with links to media and investor summary in the comments section We couldn't get the land, however, so the concept building will remain just that.
I think people got used to 50 years of Moore Law. 50 years of continuous exponential growth is exceptional and could never last.
I'm not sure why Sketch-Up wasn't faster for you (people on the SketchUp forums seem to think it runs faster on Apple Silicon, even when emulated), but Apple Silicon was a significant improvement in performance (especially per Watt) and was rewarded with a corresponding growth in market share.
I would strongly disagree with your point about hardware stagnation. There was a period in the 2000s where Intel was resting on its laurels and not innovating, but there has been a lot more competition recently. We now have a very competitive AMD. Ampere also competes with AMD and Intel with an ARM-based server processor. Amazon designed processors for server (Graviton), ML inference (Inferentia), and ML training (Trainium). Google processors a server CPU (Axion), an ML chip. And these are just the companies I can think of.
Most likely, the reason I didn't notice any improvement has something to do with Sketch-Up not improving the non-subscription version past 2021, and probably not optimizing it for Apple hardware even before that for a year or more.
The coordination - some might say collusion - between makers of chips and software that tests those chips probably exaggerates test performance improvements well beyond real world improvements when all hardware components are factored in including the OS, the 4TB hard drive, interface modules etc.
It's certainly also possible other manufacturers make better, faster, computers, but they may be more expensive niche machines that cost a lot more. Apple seems less willing to lead the market nowadays, resting on their laurals. And paying their C-Suite 10s of millions from stock buybacks.
Don't overthink it, all those processing improvements just enabled worse software to get written but still run at an acceptable speed. We long ago hit the point of computers being fast enough for the majority of applications, so now anything faster than that point just enabled us to be lazier in software.
Yes, this is the problem, particularly since code-generators, and AI, are now tasked with writing code. I started my IT career on an Atari with 64K(!) writing in BASIC, then went onto IBM PCs where a MB seemed like a ocean of RAM. Nothing tightens code like knowing you have a hard limit.
The problem is though, there is a growing subset of very high performance demanding applications, and it's not clear who's marketing to that segment in a mass market way. I long since stopped tracking the iteration of CPUs and now GPUs, and throwing RAM at the problem doesn't seem to work either. Someone needs to address this better.
> The coordination - some might say collusion - between makers of chips and software that tests those chips probably exaggerates test performance improvements
There are many open source benchmark suites (Passmark, cinebench, geekbench, etc etc) that anyone can run and upload the results to the internet. There are thousands of independent journalists and tech reviewers that have the means to buy devices and run these tests themselves. This part of the tech world is simply too open for collision to be effective.
Wonderful article. Because of my job I was able to visit a number of fabs in the US and Asia. They were both awesome and enervating. The awesome part is well documented here. The enervating part is that being inside a fab is very dehumanizing. Wearing a bunny suit removes all of your identity. It is difficult to recognize the people around you as they are all dressed white and your faces are covered. Sometimes the women wore pink bunny suits and regular workers adopted some sort of badging to be recognized. Also, the bizarre yellow lighting and high noise levels from all of the fans and pumps takes a toll on you. As fascinating as it was to be inside it was always a relief to get outside and see trees and green grass and blue sky.
Ok, I’m really curious as to how they explained that lunar phases and menstrual cycles affected semiconductor manufacturing. That sounds way too bizarre
Menstrual cycles temporarily affect the ph (acidity) of the oils and sweat on feminine skin. When I worked in a transformer plant that handled extremely fine copper wire, the women line workers had their periods charted by their supervisors, and for that part of the month had to wear finger cots (essentially miniature condoms) to isolate finger tips from the wire lest the acidity cause eventual part failures. These parts had extreme reliability requirements because they were used in things like satellites, tactical aircraft, and nuclear reactor controls. And yet were nowhere near as complicated as semiconductors.
Wow. Do humans build anything else that rivals this complexity? But given this complexity, it would seem that a semiconductor fab would be the most vulnerable facility in the world for disruption or permanently crippling by sabotage, either by outside or inside actors, or by targeted strikes by military forces. What I didn't see in the article is how many of these fabs exist, and where are they located. My guess is that it's not economical to construct low capacity fabs in the interest of distributing them across the landmass of a country for potential wartime survivability reasons. So there are probably only a handful of fabs per country.
How many fabs are there in Taiwan? Think about the risk calculations that investors and owners of Taiwanese fabs would be doing if the PRC initiated a military assault on Taiwan.
If we got into a prolonged conflict with China, would chip fabs be early, primary targets? If fabs on both sides were put out of action, that would mean the belligerents would be completely dependent upon the weapons systems each side had operational at the beginning of the war or were able to be built with chips on hand. Nothing built after the shooting started would work without chips.
Maybe we should buy all the chips China can produce in the next few years and just put them into storage?
I think military tech is typically not built with the latest and greatest fab processes, that goes into iPads, so Raytheon and BAe could still do their dirty work if the PLA has attacked the gigafabs in Taiwan. Certainly they are of interest to China from a strategic/economic perspective, but they're also something that's quite hard to steal through military action compared to the land, slaves, gold or oil that were targeted in wars past
Simply fantastic article. Started my career, out of college, in 1972 at Analog devices. My life was spent working in and out of semiconductors. Retired in 2018 from a space electronics company. Saw my IC's go in the Voyagers, GPS satellites, and wafers from 2 inches to 6 inch. 1970s processes were so primitive, it's hard to imagine todays.
Thank you for a great article
Glenn
(And a microscan of my signature is on Mars.... thanks to Insight.)
Great post, lots of information. One thing I wonder about: water. Is there any reason the water can't be recycled, given that it goes through a comprehensive purification cycle?
Working in the Tx Instrument Stafford plant outside of Houston in the 80s we were all amazed when a new 2 million dollar Perkin Elmer photo lithograph came on line.
My wife's uncle was a big shot at that plant, he and 3 other big shots left the company and built their own fabrication plant.
I have worked in the semiconductor industry for over 40 years. first as a line maintenance tech, when they had such things, and then into the construction side of things for the past several decades. The article was very good, and something that I can point people to when they ask me what I do. I have built new fabs and retrofitted older fabs. this article even covered the basic process that I learned so many decades ago, I remember when 5 layers of metal was stretching the limits. Those 50mm wafers where so cute.
Chips without dip does not go very far. The ICs in most cases need packaging to couple to substrates to complete an electronic device - unless the entire function is designed onto the Si or GaN, etc. Then what? Who and where completes the stuff afterwards?
Really enjoyed reading this article, absolutely awesome job. As a retired EE with some experience with IC design and fab (back in the 70s), I can really appreciate the engineering innovation that goes into a modern fab.
I can’t help but wonder how the engineering challenge of a modern fab compares to, say, an AP1000 nuclear plant. Seems like a nuclear plant is much, much simpler. Nuclear costs are probably around 5 times what they should be due to over wrought safety concerns.
A new fab to make 5nm devices or smaller on 300mm wafers if a near miracle. But if you want to know a system much more complex, look at the human body, biochemistry, aging, the biosphere that are much more complicated and much harder to understand comprehensively. However, faster chips, more data storage, faster telecom, better SW, AI, perhaps quantum computing at some point, could all contribute to making a greens, safer, healthier planet for 8 billion - profitably. Let's go beyond More Moore, with Semicon and Technology, photonics, EUV etc help.
Then we need more PV, Wind, other other power sources, even perhaps nuclear, to power all of the above!
Despite these incredible complications and requirements, my Apple Studio is no faster using CPU/Graphics intensive programs like Sketch-Up than my 2015-manufactured iMac I replaced it with. The new machine uses a Apple M2 Ultra with 64GB RAM, double the RAM of the old machine, but without the Intel CPU and Nvidea GPU. Maybe these old chipmakers knew something that Apple does not when it replaced their chips with a combo chip of its own, or maybe Apple sought to cut costs and sacrificed function. The new machine cost about what the old one did, despite 8 years of inflation between them, with extra cost because the monitor is not separate if you want 27", which I did (the newer iMacs only come in 24" models).
My major model is a 380MB Sketch-Up model building, but that was really close to 800MB before Sketch-Up compressed the filesize in their latest desktop version (and last; now they want a yearly subscription for the newer versions). It is noticeably slower to load, manipulate, etc., especially if other things are running. If I had known that up front, I would have just replaced the hard drive in the old machine when that died. In contrast, the 2015 iMac was much faster than the 2009 iMac it replaced, which was so slow it became unusable for my model and had to be replaced.
Apple, meanwhile, just announced a $100b stock buyback, on top of similar buybacks last year. It's clear they are not investing the same money into computers as previously, and even the endless iPhone upgrades are just evolutionary, not revolutionary.
Nothing can accelerate an industry that has decided to stagnate.
A video fly-through is here: https://bit.ly/Riverarch, with links to media and investor summary in the comments section We couldn't get the land, however, so the concept building will remain just that.
I think people got used to 50 years of Moore Law. 50 years of continuous exponential growth is exceptional and could never last.
I'm not sure why Sketch-Up wasn't faster for you (people on the SketchUp forums seem to think it runs faster on Apple Silicon, even when emulated), but Apple Silicon was a significant improvement in performance (especially per Watt) and was rewarded with a corresponding growth in market share.
I would strongly disagree with your point about hardware stagnation. There was a period in the 2000s where Intel was resting on its laurels and not innovating, but there has been a lot more competition recently. We now have a very competitive AMD. Ampere also competes with AMD and Intel with an ARM-based server processor. Amazon designed processors for server (Graviton), ML inference (Inferentia), and ML training (Trainium). Google processors a server CPU (Axion), an ML chip. And these are just the companies I can think of.
Most likely, the reason I didn't notice any improvement has something to do with Sketch-Up not improving the non-subscription version past 2021, and probably not optimizing it for Apple hardware even before that for a year or more.
The coordination - some might say collusion - between makers of chips and software that tests those chips probably exaggerates test performance improvements well beyond real world improvements when all hardware components are factored in including the OS, the 4TB hard drive, interface modules etc.
It's certainly also possible other manufacturers make better, faster, computers, but they may be more expensive niche machines that cost a lot more. Apple seems less willing to lead the market nowadays, resting on their laurals. And paying their C-Suite 10s of millions from stock buybacks.
Don't overthink it, all those processing improvements just enabled worse software to get written but still run at an acceptable speed. We long ago hit the point of computers being fast enough for the majority of applications, so now anything faster than that point just enabled us to be lazier in software.
Yes, this is the problem, particularly since code-generators, and AI, are now tasked with writing code. I started my IT career on an Atari with 64K(!) writing in BASIC, then went onto IBM PCs where a MB seemed like a ocean of RAM. Nothing tightens code like knowing you have a hard limit.
The problem is though, there is a growing subset of very high performance demanding applications, and it's not clear who's marketing to that segment in a mass market way. I long since stopped tracking the iteration of CPUs and now GPUs, and throwing RAM at the problem doesn't seem to work either. Someone needs to address this better.
> The coordination - some might say collusion - between makers of chips and software that tests those chips probably exaggerates test performance improvements
There are many open source benchmark suites (Passmark, cinebench, geekbench, etc etc) that anyone can run and upload the results to the internet. There are thousands of independent journalists and tech reviewers that have the means to buy devices and run these tests themselves. This part of the tech world is simply too open for collision to be effective.
Wonderful article. Because of my job I was able to visit a number of fabs in the US and Asia. They were both awesome and enervating. The awesome part is well documented here. The enervating part is that being inside a fab is very dehumanizing. Wearing a bunny suit removes all of your identity. It is difficult to recognize the people around you as they are all dressed white and your faces are covered. Sometimes the women wore pink bunny suits and regular workers adopted some sort of badging to be recognized. Also, the bizarre yellow lighting and high noise levels from all of the fans and pumps takes a toll on you. As fascinating as it was to be inside it was always a relief to get outside and see trees and green grass and blue sky.
Thanks for writing this wonderful article. I am taking a paid membership just to get the further reading list.
Ok, I’m really curious as to how they explained that lunar phases and menstrual cycles affected semiconductor manufacturing. That sounds way too bizarre
Menstrual cycles temporarily affect the ph (acidity) of the oils and sweat on feminine skin. When I worked in a transformer plant that handled extremely fine copper wire, the women line workers had their periods charted by their supervisors, and for that part of the month had to wear finger cots (essentially miniature condoms) to isolate finger tips from the wire lest the acidity cause eventual part failures. These parts had extreme reliability requirements because they were used in things like satellites, tactical aircraft, and nuclear reactor controls. And yet were nowhere near as complicated as semiconductors.
Do they serve beans in the cafeteria?
Well by lunar phases I assume is meant that the tidal forces exerted by the moon can affect the process.
Wow. Do humans build anything else that rivals this complexity? But given this complexity, it would seem that a semiconductor fab would be the most vulnerable facility in the world for disruption or permanently crippling by sabotage, either by outside or inside actors, or by targeted strikes by military forces. What I didn't see in the article is how many of these fabs exist, and where are they located. My guess is that it's not economical to construct low capacity fabs in the interest of distributing them across the landmass of a country for potential wartime survivability reasons. So there are probably only a handful of fabs per country.
How many fabs are there in Taiwan? Think about the risk calculations that investors and owners of Taiwanese fabs would be doing if the PRC initiated a military assault on Taiwan.
If we got into a prolonged conflict with China, would chip fabs be early, primary targets? If fabs on both sides were put out of action, that would mean the belligerents would be completely dependent upon the weapons systems each side had operational at the beginning of the war or were able to be built with chips on hand. Nothing built after the shooting started would work without chips.
Maybe we should buy all the chips China can produce in the next few years and just put them into storage?
If all you want is low tech that would be fine.
Musk bought an old plant. He didn't need the newest, smallest, fastest chip.
I think military tech is typically not built with the latest and greatest fab processes, that goes into iPads, so Raytheon and BAe could still do their dirty work if the PLA has attacked the gigafabs in Taiwan. Certainly they are of interest to China from a strategic/economic perspective, but they're also something that's quite hard to steal through military action compared to the land, slaves, gold or oil that were targeted in wars past
Simply fantastic article. Started my career, out of college, in 1972 at Analog devices. My life was spent working in and out of semiconductors. Retired in 2018 from a space electronics company. Saw my IC's go in the Voyagers, GPS satellites, and wafers from 2 inches to 6 inch. 1970s processes were so primitive, it's hard to imagine todays.
Thank you for a great article
Glenn
(And a microscan of my signature is on Mars.... thanks to Insight.)
Great post, lots of information. One thing I wonder about: water. Is there any reason the water can't be recycled, given that it goes through a comprehensive purification cycle?
Working in the Tx Instrument Stafford plant outside of Houston in the 80s we were all amazed when a new 2 million dollar Perkin Elmer photo lithograph came on line.
My wife's uncle was a big shot at that plant, he and 3 other big shots left the company and built their own fabrication plant.
They weren't rich men, not even millionaires.
Amazing article on fabs, thanks.
I have worked in the semiconductor industry for over 40 years. first as a line maintenance tech, when they had such things, and then into the construction side of things for the past several decades. The article was very good, and something that I can point people to when they ask me what I do. I have built new fabs and retrofitted older fabs. this article even covered the basic process that I learned so many decades ago, I remember when 5 layers of metal was stretching the limits. Those 50mm wafers where so cute.
Being a part of this amazing semiconductors universe, really enjoyed reading your detailed article.
Absolutely fantastic article, comprehensively answers the question "how the heck can something cost so much?!"
I suspect the answers for nuclear plants and F35 jets are not quite as satisfying...
Chips without dip does not go very far. The ICs in most cases need packaging to couple to substrates to complete an electronic device - unless the entire function is designed onto the Si or GaN, etc. Then what? Who and where completes the stuff afterwards?
Hey man! Write about semiconductors for a living, and you did an outstanding job here. :)
https://open.substack.com/chat/posts/39658bce-569d-47e1-9327-631af16f2f69?utm_source=share&utm_medium=android&r=19xmh1