If this was making things like cars or toilet paper, where the design and manufacturing methods are well understood, that’s exactly what would happen.
Making computer chips was well understood. You take a slice of single-crystal silicon, dope its surface with something, use a photo mask and a light to etch off the doping where you don’t want it. Repeat that process a few dozen or a few hundred times to etch transistors into the silicon, then slice it up into chips and package them.
Machines that do this are readily available.
The problem is, we made those transistors smaller. And smaller. And smaller. And smaller. And smaller.
Then we ran into problems like how to etch features on the silicon that are smaller than the wavelength of light. And we found solutions to that. And so we made the transistors smaller. And smaller. And smaller.
So now to create a current-gen computer chip, you need a process called EUV- Extreme Ultra Violet. Red light has a longer wavelength than blue light, so when you want smaller wavelength to etch smaller features into the chip you eventually go from red to blue to ultraviolet and eventually to extreme ultraviolet. Problem THERE is there’s no EUV bulbs available, and if there were they’d be useless because atmospheric air absorbs EUV, and even if you do it in a vacuum glass absorbs EUV too so lenses don’t work. So you can only manipulate this light with mirrors, which have to be ground to an absolutely insane level of precision.
The resulting machine is quite impressive. You have a giant cavity kept in perfect vacuum. In one side you have an EUV source- that’s a little machine that dispenses a tiny 3-micron droplet of molten tin. As the droplet falls it’s hit by a laser to blast it into a pancake-like shape, then by another much bigger laser that vaporizes it. In the process of vaporizing it releases EUV light. So by blasting 50,000 droplets per second, you have a mostly continuous EUV light source. This light is reflected by shaped mirrors carved into the surfaces of the vacuum chamber that reflect and focus the EUV into a linear beam. Below that (also in vacuum) you have the mask and the silicon wafer, and by moving them back and forth under the beam you etch features ‘smaller than light’ into the chip.
The result is a chip with current pathways less than 100 silicon atoms wide. And if you want to make current-gen computer memory, that’s the only way we’ve got to build it.
EUV machines are pretty much only made by one company, ASML. They’re the size of a double decker bus, they cost a fortune ($200MM+), and they’re in insanely high demand. Like I’ve heard of a guy getting hired by a company for over million a year simply because he’s friends with a purchasing manager at ASML and might be able to get the company that hired him a build slot.
And it’s not just buy the machine and hit ‘start’, there’s a ton of other stuff involved. You need machines to make the photo masks, you need an ultra-clean cleanroom, you need a robotic facility where cassettes of wafers can be whisked from machine to machine with no exposure to even clean room air. And this is all highly specialized stuff, you don’t just call 1800-fab-4you and place an order.
Bottom line- even for a company that already has experience in current-gen chipmaking, setting up a fab like this costs $15-20 billion. And it isn’t just ‘sign a check and come back tomorrow’, the process of building a fab from the project being approved to the first wafer coming off the line is 2-3 years minimum.
Now here’s the bigger problem- semiconductors are always a cyclical market, or at least always have been. Demand (and thus prices) goes up, demand/prices come down. So if you invest $20 billion when prices are high, then in 2-3 years when the cycle is at its low nobody’s gonna be buying your output. And of course, adding more chips to the market will affect market prices (supply and demand). So companies that are building fabs have to look 5-10 years ahead to determine if they’ll get ROI on a fab before they build.
And that brings us to the next issue- with AI, we’re in uncharted territory. The computing market has been pretty well understood since the early 90s. There’s demand for PCs and laptops and servers and gadgets, and it goes up and down and new products come out that changes the mix of what’s ordered, but the cycle more or less continues. Up and down.
Then AI happens. And suddenly we have near-instant, unheard-of levels of demand. And it’s all for current-gen top-shelf stuff- HBM (high bandwidth memory) and GPUs and specialty silicon like NPUs.
Now there are more fabs being built. But it’s also starting to be better understood that AI is a bubble, which almost certainly will pop. So if you’re a DRAM maker and you spend $40 billion building a fleet of new fabs and then the bubble pops, you’re gonna be fucked. That’s why you don’t see everybody+dog diving into the DRAM market face first.
All true but it also assumes valve isn’t thinking post gaben
Building a fab now and getting the experience and expertise it needs to build next gen, such as abandoning silicon like the one article I saw said might be the direction things go in might be in valves orbit. Froml all accounts valve is highly profitable and could potentially weather the storm of upfront capital and growing/learning pains.
Also as I understand it valve doesn’t hire developers that don’t have a minimum of 15 years in industry related employment so if valve maintains its current hiring standards (and I’m not completely wrong or it has already changed), they VERY easily could poach all the experience they would need to since they would have the time. Might be a very exciting and enthralling opportunity to work for a gamer/pc focused fab.
I mean I doubt it but if they were thinking long term it would given them a comfortable time frame to do it in.
If turns out they already have the capital to self finance it all that wouldn’t entirely surprise me they could.
Just to add, since I didn’t see you mention it: even when you build a fab from scratch and know what to fabricate you need to find and hire people with the “know-how” on how to set the processes up.
In some ways yes. Older designs on older nodes use older equipment- less precise, easier to build. There’s actually a lot of PC builders who are building machines on DDR4 mainboards because the memory is more affordable. A couple of DDR4 mainboards have been ‘un-discontinued’ as a result.
For a desktop or laptop PC, the difference of DDR4 to DDR5 often doesn’t make a huge difference. So you get a last-gen chipset and RAM and you get a decent machine without paying $thousands extra.
The problem is it’s not quite so easy.
If this was making things like cars or toilet paper, where the design and manufacturing methods are well understood, that’s exactly what would happen.
Making computer chips was well understood. You take a slice of single-crystal silicon, dope its surface with something, use a photo mask and a light to etch off the doping where you don’t want it. Repeat that process a few dozen or a few hundred times to etch transistors into the silicon, then slice it up into chips and package them.
Machines that do this are readily available.
The problem is, we made those transistors smaller. And smaller. And smaller. And smaller. And smaller.
Then we ran into problems like how to etch features on the silicon that are smaller than the wavelength of light. And we found solutions to that. And so we made the transistors smaller. And smaller. And smaller.
So now to create a current-gen computer chip, you need a process called EUV- Extreme Ultra Violet. Red light has a longer wavelength than blue light, so when you want smaller wavelength to etch smaller features into the chip you eventually go from red to blue to ultraviolet and eventually to extreme ultraviolet. Problem THERE is there’s no EUV bulbs available, and if there were they’d be useless because atmospheric air absorbs EUV, and even if you do it in a vacuum glass absorbs EUV too so lenses don’t work. So you can only manipulate this light with mirrors, which have to be ground to an absolutely insane level of precision.
The resulting machine is quite impressive. You have a giant cavity kept in perfect vacuum. In one side you have an EUV source- that’s a little machine that dispenses a tiny 3-micron droplet of molten tin. As the droplet falls it’s hit by a laser to blast it into a pancake-like shape, then by another much bigger laser that vaporizes it. In the process of vaporizing it releases EUV light. So by blasting 50,000 droplets per second, you have a mostly continuous EUV light source. This light is reflected by shaped mirrors carved into the surfaces of the vacuum chamber that reflect and focus the EUV into a linear beam. Below that (also in vacuum) you have the mask and the silicon wafer, and by moving them back and forth under the beam you etch features ‘smaller than light’ into the chip.
The result is a chip with current pathways less than 100 silicon atoms wide. And if you want to make current-gen computer memory, that’s the only way we’ve got to build it.
EUV machines are pretty much only made by one company, ASML. They’re the size of a double decker bus, they cost a fortune ($200MM+), and they’re in insanely high demand. Like I’ve heard of a guy getting hired by a company for over million a year simply because he’s friends with a purchasing manager at ASML and might be able to get the company that hired him a build slot.
And it’s not just buy the machine and hit ‘start’, there’s a ton of other stuff involved. You need machines to make the photo masks, you need an ultra-clean cleanroom, you need a robotic facility where cassettes of wafers can be whisked from machine to machine with no exposure to even clean room air. And this is all highly specialized stuff, you don’t just call 1800-fab-4you and place an order.
Bottom line- even for a company that already has experience in current-gen chipmaking, setting up a fab like this costs $15-20 billion. And it isn’t just ‘sign a check and come back tomorrow’, the process of building a fab from the project being approved to the first wafer coming off the line is 2-3 years minimum.
Now here’s the bigger problem- semiconductors are always a cyclical market, or at least always have been. Demand (and thus prices) goes up, demand/prices come down. So if you invest $20 billion when prices are high, then in 2-3 years when the cycle is at its low nobody’s gonna be buying your output. And of course, adding more chips to the market will affect market prices (supply and demand). So companies that are building fabs have to look 5-10 years ahead to determine if they’ll get ROI on a fab before they build.
And that brings us to the next issue- with AI, we’re in uncharted territory. The computing market has been pretty well understood since the early 90s. There’s demand for PCs and laptops and servers and gadgets, and it goes up and down and new products come out that changes the mix of what’s ordered, but the cycle more or less continues. Up and down.
Then AI happens. And suddenly we have near-instant, unheard-of levels of demand. And it’s all for current-gen top-shelf stuff- HBM (high bandwidth memory) and GPUs and specialty silicon like NPUs.
Now there are more fabs being built. But it’s also starting to be better understood that AI is a bubble, which almost certainly will pop. So if you’re a DRAM maker and you spend $40 billion building a fleet of new fabs and then the bubble pops, you’re gonna be fucked. That’s why you don’t see everybody+dog diving into the DRAM market face first.
All true but it also assumes valve isn’t thinking post gaben
Building a fab now and getting the experience and expertise it needs to build next gen, such as abandoning silicon like the one article I saw said might be the direction things go in might be in valves orbit. Froml all accounts valve is highly profitable and could potentially weather the storm of upfront capital and growing/learning pains.
Also as I understand it valve doesn’t hire developers that don’t have a minimum of 15 years in industry related employment so if valve maintains its current hiring standards (and I’m not completely wrong or it has already changed), they VERY easily could poach all the experience they would need to since they would have the time. Might be a very exciting and enthralling opportunity to work for a gamer/pc focused fab.
I mean I doubt it but if they were thinking long term it would given them a comfortable time frame to do it in.
If turns out they already have the capital to self finance it all that wouldn’t entirely surprise me they could.
That’s probably the best comment Ive read so far on all of Lemmy, thanks!
Most welcome :)
Just to add, since I didn’t see you mention it: even when you build a fab from scratch and know what to fabricate you need to find and hire people with the “know-how” on how to set the processes up.
This was a very interesting read, thank you!
Most welcome :)
Would DDR4 or DDR3 be easier to produce?
In some ways yes. Older designs on older nodes use older equipment- less precise, easier to build. There’s actually a lot of PC builders who are building machines on DDR4 mainboards because the memory is more affordable. A couple of DDR4 mainboards have been ‘un-discontinued’ as a result.
For a desktop or laptop PC, the difference of DDR4 to DDR5 often doesn’t make a huge difference. So you get a last-gen chipset and RAM and you get a decent machine without paying $thousands extra.