IBM 2 NM chipmaking breakthrough?

Sal1

https://www.msn.com/en-us/news/technology/ibm-nanosheets-promise-better-speed-and-battery-life-for-next-gen-chips/ar-BB1gq0Gg

"The new manufacturing technology, featuring components called nanosheets, increases chip performance 45% or reduces power consumption by 75% compared with the one used to make IBM server chips or Apple's iPhone chips, IBM said Thursday. The company expects the technology to arrive in processors in 2024 or 2025, two generations beyond the most advanced processes currently used by today's manufacturing leader Taiwan Semiconductor Manufacturing, or TSMC."

"Another complication for IBM will be proving its technology works outside of a lab in high-volume manufacturing, where costs and consistency are crucial."

I think that is critical.

olepi

TSMC has a 5nm process now, and is working on 3nm. It will be interesting to see if IBM can actually make large production runs in 2nm.

"Our first commercialized 7 nm processor, based on IBM Research’s 2015 breakthrough, will debut later this year in IBM POWER10 processor, for example." -- IBM

That doesn't sound very promising. TSMC has had 7nm in production for several years already, and has 5nm ready, is working on 3nm, and says 2nm is "on track".

If IBM is just now producing 7nm ....

Slapshot1188

The more the better is what I say.  We need diversified sources of chips.

Vrika

Mars_OMG said:

in my opinion IBM will not be able to compete with AMD's price per chip, which is the true bottom line when it comes to server implementation.

Would also say IBM as a product is a tailor made solution for the "biggest" companies in the world, not so much a generic solution to 90% of server market shares.

As Intel and AMd goes, this will hurt Intel more as they continue to lose more and more market shares in server sales.

This news is not really about how IBM might or might not compete against AMD, since it's tech for foundries and AMD does not have any tech for foundries or their own foundries.

It's a tech that AMD could potentially one day end up using if IBM licenses it to some foundry and then AMD buys manufacturing from that foundry, rather than something that'd bring them into competition.

For Intel it might end up hurting them if IBM licenses it to one of the competing foundries. But it's also not impossible (though not very likely) that Intel could license this tech for their own foundries if IBM gets it working.

Vrika

olepi said:

TSMC has a 5nm process now, and is working on 3nm. It will be interesting to see if IBM can actually make large production runs in 2nm.

"Our first commercialized 7 nm processor, based on IBM Research’s 2015 breakthrough, will debut later this year in IBM POWER10 processor, for example." -- IBM

That doesn't sound very promising. TSMC has had 7nm in production for several years already, and has 5nm ready, is working on 3nm, and says 2nm is "on track".

If IBM is just now producing 7nm ....

IBM is not actually producing any chips. Their 7nm processors are produced by Samsung. Even if IBM develops the tech, they do not have their own foundries.

Quizzical

IBM sold their fabs to the Emirate of Abu Dhabi several years ago.  This was shortly after AMD did the same thing.  That's where Global Foundries came from.  And Global Foundries has abandoned the cutting edge to focus on continuing to produce chips for older process nodes.  I'm not sure why IBM still does research like that.

Process node names aren't really that meaningful anymore.  Just because two different fabs both call a process node "7 nm" doesn't mean that they're the same thing or even kind of similar.  One could easily be a lot better than another.

Cuddleheart

Isn't the rub on this process the fact that they can stack the gates or something like that?

Sal1

Quizzical said:

IBM sold their fabs to the Emirate of Abu Dhabi several years ago.  This was shortly after AMD did the same thing.  That's where Global Foundries came from.  And Global Foundries has abandoned the cutting edge to focus on continuing to produce chips for older process nodes.  I'm not sure why IBM still does research like that.

Process node names aren't really that meaningful anymore.  Just because two different fabs both call a process node "7 nm" doesn't mean that they're the same thing or even kind of similar.  One could easily be a lot better than another.

Sold their fabs? What does that mean?

Ridelynn

Sal1 said:

Quizzical said:

IBM sold their fabs to the Emirate of Abu Dhabi several years ago.  This was shortly after AMD did the same thing.  That's where Global Foundries came from.  And Global Foundries has abandoned the cutting edge to focus on continuing to produce chips for older process nodes.  I'm not sure why IBM still does research like that.

Process node names aren't really that meaningful anymore.  Just because two different fabs both call a process node "7 nm" doesn't mean that they're the same thing or even kind of similar.  One could easily be a lot better than another.

Sold their fabs? What does that mean?

IBM doesn't make chips anymore. They may have developed a 2nm process, but they won't make anything with it. The best they can do is patent the process and license it to a foundry that does manufacture chips. It was sold to the same company that AMD sold their manufacturing to (GloFo, as Quiz mentioned)

IBM ~used~ to make their own chips, probably most people would recognize the PowerPC used in XBox 360 and Wii -- they were different chips but both made by IBM, but they sold off that part of the business back in 2014.

Sal1

Torval said:

While the nanometer naming convention is somewhat arbitrary the transistor density of this process is well beyond what TSMC, Intel, or Samsung are currently offering. This isn't just a marketing blurb. It is significant progress in hardware architecture.

If anyone is interested this showed up on Hacker News a few days ago. Here is the thread discusion: https://news.ycombinator.com/item?id=27061130

Here is the posted article for that discussion: https://www.anandtech.com/show/16656/ibm-creates-first-2nm-chip

There is some great info packed into the Anand article. The transistor density is 333 million per square millimeter.

If you're wondering why IBM does this sort of research, Dr. Cutress who wrote the article says, "Users might be wondering why we’re hearing that IBM is the first to a 2nm chip. IBM is one of the world’s leading research centers on future semiconductor technology, and despite not having a foundry offering of their own, IBM develops IP in collaboration with others for their manufacturing facilities."

Thank you Torval. But after reading that it sounds like Xerox PARC all over again. lol

Anyone that is young do your research and find out about how Xerox should have owned the PC industry. Just my opinion.

Sal1

Ridelynn said:

Sal1 said:

Quizzical said:

IBM sold their fabs to the Emirate of Abu Dhabi several years ago.  This was shortly after AMD did the same thing.  That's where Global Foundries came from.  And Global Foundries has abandoned the cutting edge to focus on continuing to produce chips for older process nodes.  I'm not sure why IBM still does research like that.

Process node names aren't really that meaningful anymore.  Just because two different fabs both call a process node "7 nm" doesn't mean that they're the same thing or even kind of similar.  One could easily be a lot better than another.

Sold their fabs? What does that mean?

IBM doesn't make chips anymore. They may have developed a 2nm process, but they won't make anything with it. The best they can do is patent the process and license it to a foundry that does manufacture chips. It was sold to the same company that AMD sold their manufacturing to (GloFo, as Quiz mentioned)

IBM ~used~ to make their own chips, probably most people would recognize the PowerPC used in XBox 360 and Wii -- they were different chips but both made by IBM, but they sold off that part of the business back in 2014.

The best they can do? How about manufacturing the chips in the USA like you used to do?

Quizzical

Sal1 said:

Ridelynn said:

Sal1 said:

Quizzical said:

IBM sold their fabs to the Emirate of Abu Dhabi several years ago.  This was shortly after AMD did the same thing.  That's where Global Foundries came from.  And Global Foundries has abandoned the cutting edge to focus on continuing to produce chips for older process nodes.  I'm not sure why IBM still does research like that.

Process node names aren't really that meaningful anymore.  Just because two different fabs both call a process node "7 nm" doesn't mean that they're the same thing or even kind of similar.  One could easily be a lot better than another.

Sold their fabs? What does that mean?

IBM doesn't make chips anymore. They may have developed a 2nm process, but they won't make anything with it. The best they can do is patent the process and license it to a foundry that does manufacture chips. It was sold to the same company that AMD sold their manufacturing to (GloFo, as Quiz mentioned)

IBM ~used~ to make their own chips, probably most people would recognize the PowerPC used in XBox 360 and Wii -- they were different chips but both made by IBM, but they sold off that part of the business back in 2014.

The best they can do? How about manufacturing the chips in the USA like you used to do?

Every new process node costs more than the previous.  There used to be dozens of companies that had their own fabs to produce their own chips.  As the costs of keeping your fabs up to date increased, one by one, they decided that they no longer had the volume necessary to justify having their own fabs.  Instead, it would be cheaper to pay someone else to build their chips.  That effectively allowed them to share the costs of keeping the fabs up to date with other companies.

Today, there are only three companies that aspire to build cutting-edge fabs:  TSMC, Samsung, and Intel.  Well, maybe you could say that SMIC has the aspiration, but they're years behind their foreign competitors and mostly rely on being subsidized by their government to keep them in business.  TSMC and Samsung each sell foundry capacity to many other customers, and Intel plans to do the same very soon.

Ridelynn

It's Moore's Second Law of Computing.

While the first law states that transistor count will double every 2 years (which is loosely interpreted as speed, more or less accurate to some extent), Moore stated (or Arthur Rock, depending on which story you believe) in his second law that cost of the fab plant to produce those chips would double every 4 years.

So yeah, here we are what, 60 years after Moore/Rock saying that... that's a lot of cost doubling that's gone on.

Rungar

The first law of computers is that all critical computer infrastructure must have ~13% combined shares owned by Vanguard and Blackrock. 

Quizzical

Torval said:

Ridelynn said:

It's Moore's Second Law of Computing.

While the first law states that transistor count will double every 2 years (which is loosely interpreted as speed, more or less accurate to some extent), Moore stated (or Arthur Rock, depending on which story you believe) in his second law that cost of the fab plant to produce those chips would double every 4 years.

So yeah, here we are what, 60 years after Moore/Rock saying that... that's a lot of cost doubling that's gone on.

Another reason this tech is interesting is how it uses stacking to achieve greater density. TSMC and Samsung are also moving to a stacking technology according to the Cutress in the Anand article. The idea is similar to NAND stacking in SSDs.

It seems like the sort of stacking being employed here would be somewhat limited in that the transistors couldn't be stacked in ever greater numbers due to thermal limits. The more stacks the harder it will be to keep the chip cool.

GAAFETs are perhaps a sort of stacking.  You could already think of the metal layers in a logic chip as being stacked on top of each other, and that's been going on for decades.  But you can't just stack high-powered logic chips on top of each other the same way that you can with DRAM and NAND.  You need a good way to cool whatever chip is in the middle, and putting very hot chips on top of and below it is not going to work.