IBM Research Alliance Has Figured Out How To Make 5nm Chips (cnet.com)

IBM, GlobalFoundries, and Samsung said Monday that they have found a way to make thinner transistors, which should enable them to pack 30 billion switches onto a microprocessor chip the size of a fingernail. The tech industry has been fueled for decades by the ability of chipmakers to shoehorn ever smaller, faster transistors into the chips that power laptops, servers, and mobile devices. But industry watchers have worried lately that technology was pushing the limits of Moore's Law -- a prediction made by Intel co-founder Gordon Moore in 1965 that computing power would double every two years as chips got more densely packed. From a report: Today's chips are built with transistors whose dimensions measure 10 nanometers, which means about 1,000 fit end-to-end across the diameter of a human hair. The next generation will shrink that dimension to 7nm, and the IBM-Samsung development goes one generation beyond that to 5nm. That means transistors can be packed four times as densely on a chip compared with today's technology. "A nanosheet-based 5nm chip will deliver performance and power, together with density," said Huiming Bu, IBM's director of silicon integration and device research. Take all those numbers with a nanograin of salt, though, because chipmakers no longer agree on what exactly they're measuring about transistors. And there's also a long road between this research announcement and actual commercial manufacturing. IBM believes this new process won't cost any more than chips with today's transistor designs, but its approach requires an expensive shift that chipmakers have put off for years: the use of extreme ultraviolet light to etch chip features onto silicon wafers.

Comments

Moore's law

[Whatanut]

I'll just leave this here.

https://en.wikipedia.org/wiki/...

Re:Moore's law

[Anonymous Coward]

Since Whatanut doesn't spell it out. Moore's law is an observation that number of transisters in a dense integrated circuit doubles approximately every two years. This is a tech site, we should get it right - shame on msmash. Really, why go to the trouble to name Moore's first name and the year, and get what he actually said wrong.

OK. Enough discussion on the inaccuracy. What do yall think of their 5nm claim?

Re:Moore's law

[NoNonAlphaCharsHere]

, I know. Moore's Law (observation, actually) is that transistor count (approximately) doubles every 2 years, not "power" (current times voltage?), or "processing power" (Whetstones perhaps?), and certainly not "speed" (GHz times bus width?). This process seems to obey the transistor count rule, but with heat already being the problem it is, it's hard to say what quadrupling the density actually buys you.

Re:Moore's law

[Khyber]

"What do yall think of their 5nm claim?"

They just about hit the limits of atomic transistors, if this is true and the feature size they refer to is not the trace but the transistor itself.

From that, I expect to see at least an IPC improvement of 2 or 3 per thread given the shitty bloated coding everyone does now days, which should put us back on par with how things worked when we were on server-class P3 dies and didn't need these extra bullshit instructions.

Re:Moore's law

[Anonymous Coward]

When density goes up the electrical current goes down, or you're frying it. Chips get smaller, more powerful AND more efficient all at once, or none of the three really.

Re:Moore's law

[Archangel Michael]

Density also shortens pathways, so that the clock speed may not change, but the distance cut increases efficiency of design.

I remember seeing a demo by someone (Grace Hopper??) holding up a piece of wire, and saying "This is a microsecond", the distance speed of light travels in that time frame. It was quite illuminating to me at the time (35 years ago)

Re:Moore's law

[swillden]

I remember seeing a demo by someone (Grace Hopper??) holding up a piece of wire, and saying "This is a microsecond"

You mean nanosecond, not microsecond. Light travels about 300 meters in a microsecond, which would require a spool of wire. Light travels approximately 30 cm in a nanosecond.

Re:Moore's law

[CanadianMacFan]

You can also add new things to the chip. There are plenty of things that used to be off of the main chip and on the motherboard 20 or 30 years ago. But as space became available on the chip they were moved on. This prevented the use of the bus on the motherboard and greatly sped up things. You can also add more registers, memory, and processing units (arithmetic for example).

Everything is electrons moving around. If you can move everything closer you shorten the trip an electron has to take. Imagine if all of the city blocks where reduced by 10% with one corner anchored but you still walked the same pace. You would get where you were going faster with the same amount of energy.

Re:Moore's law

[CanadianMacFan]

But it's easier to build a faster processor than it is to retrain all of the shitty programmers writing bloated code.

Re:Moore's law

[HiThere]

Not clear. Electricity travels through wires significantly slower than light travels through a vacuum. OTOH, you can't give a generic speed for a generic wire, because it's different for different wires at different speeds. (And if you coil the wire it's worse, but she clearly wasn't coiling it.)

But it would have needed to be a long piece of wire.

Re:Moore's law

[ChrisMaple]

Increasing instructions per clock is already extremely difficult and hardware intensive. I doubt that we'll see a doubling of IPC in a decade, and it may require compilers to optimize code for high IPC.

Branches are a major consideration in increasing IPC. In order to prevent stalls caused by erroneous branch decisions, both paths may be speculatively executed as a new thread, and each branch may in turn hit another branch and in turn require a new speculative thread, all of which must be executed simultaneously to increase IPC. The hardware has to know when to abort a speculative thread, has to have the resources to execute it, and has to have very high cache bandwidth to fetch instructions and data for each speculative thread (or main memory if not in cache.) In short, increased IPC makes literally exponential demands on hardware.

I think that the difficulties involved in increasing IPC beyond the current state of the art are already into the range where the tradeoffs involved in better IPC mean that other aspects of hardware may have a better return on effort.

Re:Moore's law

[Nehmo]

Density also shortens pathways, so that the clock speed may not change, but the distance cut increases efficiency of design.

I remember seeing a demo by someone (Grace Hopper??) holding up a piece of wire, and saying "This is a microsecond", the distance speed of light travels in that time frame. It was quite illuminating to me at the time (35 years ago)

Nanosecond, that is 10^-9 second is what you mean. She was giving an idea of the speed of light in a vacuum is about a foot per nanosecond. The electrical wavefront velocity in a copper wire is somewhat slower. https://en.wikipedia.org/wiki/... or https://www.quora.com/What-is-...

Basically, Hopper was off by about an inch. The wavefront only travels 11 inches in a nanosecond.

In terms of the conductive paths on an integrated circuit, I will defer to a more knowledgeable brain.

BTW, If an inch were an AU (distance from the earth to the sun), a mile would be a lightyear.

If you stacked 100,000 cans of Campbell's soup, you'd interfere with airliners in flight.

Re:Moore's law

[swillden]

Not clear. Electricity travels through wires significantly slower than light travels through a vacuum.

True, though not really relevant for the demonstration in question (which I found on YouTube https://www.youtube.com/watch?...). Grace Hopper actually clarified that she was talking about the maximum possible distance that could be traveled in a nanosecond, so she used the speed of light in a vacuum. As it turns out, she also did show a microsecond length, in largish coil. She apparently handed out "nanoseconds" to the students she was speaking to.

Re:Moore's law

[UnknownSoldier]

> What do yall think of their 5nm claim?

Considering Silicon can't scale past ~5 GHz, whereas GaAs can scale to 500+ GHz I see this as a stop-gap solution. Sure, 5 nm will allow for more cores on a die, but single-threaded performance is STILL a bottleneck. No one seems to looking at the long game -- bio-computing -- due to the Billions invested in existing infrastructure.

At some point we need to ditch Silicon -- this is just keeping it on its dying (performance) legs.

Re:Moore's law

[tlhIngan]

This process seems to obey the transistor count rule, but with heat already being the problem it is, it's hard to say what quadrupling the density actually buys you.

Really, you cannot imagine there are any chips in your computer or smartphone today that will not benefit from a quadrupling of density?

There are many transistor-limited devices out there - where the ability to stick more transistors in a smaller area is a net plus (or more transistors in the same area).

I'm talking about memory. And memory in the general sense, at that. DRAM technology has 1 transistor per bit. Quadrupling it means you can either have 4 times the RAM for the same space, or the same RAM in much less space (with lowered power consumption or faster speeds).

And yes, it applies to storage memory, too, for it means either cheaper SSDs (the smaller the area used the cheaper the chip), more capacious SSDs (no, we won't reach 10TB in a reasonable price yet, but we can have lots more storage available).

No, general random logic (e.g., CPUs and other logic devices) will not benefit - the density of transistors in a random logic device is determined less by process technology and more by wire density - the amount of wiring required for all those transistors is dictating how close we can put them together. The only part where this isn't true is the caches, and that's where most of the transistors lie in a modern processor - the density is so high that the area used by caches can be small yet contain 80% of the transistors yet take only 30% of the space.

In fact, a modern random logic device has so low transistor density that most places will fab in tons of extra transistors and logic blocks that are not connected to active circuits. They are not there for use, but for bug fixing - minor bugs can be fixed by changing a few metal layers only, while major bugs may require re-laying out all the transistors again. The stepping often tells you what changed. Going from A0 to A1, usually means a metal layer rework (using those spare transistors laid out during design). Going from A0 to B0 means a full mask set including base transistor layer changes.

small chips

[Anonymous Coward]

small chips require small hands, you do the math!

IBM doesn't have its own fab

[Anonymous Coward]

Samsung and GF are competitors now.

Hillary lost

[Anonymous Coward]

Get over it.

Making chips is one thing

[Anonymous Coward]

But how many errors do they have statistically per year?

Switch to picometers

[SPopulisQR]

We need to switch from nanometers to picometers. 5 nm is 5000 picometers. The diameter of silicon atom is 210 picometers, thus 5nm will be equal to approx 24 silicon atom diameters, which will provide a valuable perspective. I do understand that measuring lakes in olympic swiming pools will be used to compare measuring transistors with the silicon atom radius, however in this situation there is a limit on how small transistors can exist in practice.

Re:Switch to picometers

[aicrules]

It really is pretty exciting how close to atomic our transistors are becoming. We could conceivably create transistors that are smaller than some atoms. Regardless of what unit of measurement is used, it's really hard to understand the concept of how small that is.

Re:Switch to picometers

[Anonymous Coward]

Yes but keep in mind that this number isn't the size of the transistor, it is the size of the transistor features and that is where the debate in the summary comes in, which feature are you measuring.

Re:Switch to picometers

[Dutch Gun]

We've done it before, in the switch from micrometers to nanometers. The 80386 processors were 1um processes.

To be honest, though, I don't remember anyone talking about process size until 90nm CPUs. Previously to that, everyone measured CPU performance by processor clock speed. So, I guess we needed a new performance metric to talk about (along with multi-core), since speeds more or less stalled.

Re:Switch to picometers

[ChrisMaple]

I prefer Angstroms, because that's closer to the dimemsions of atoms.

Re:Switch to picometers

[mentil]

I recall enthusiast sites talking about the .130 micrometer process, it wasn't often put in terms of nanometers until the 90nm process.

Re:Switch to picometers

[Anonymous Coward]

Nanometers is fine as a unit. A silicon atom is 0.21 of a nm, a perfectly sensible understandable unit. Realistically can we really get below 1nm feature size, that is only ~16-20 atoms square. It looks to me like 5nm is close to the end for silicon.

Re:Switch to picometers

[Dutch Gun]

That could be. Maybe it's just today that we frame the measurements in such a way with the benefit of hindsight.

EUV

[Anonymous Coward]

But isn't EUV just round the corner with only few issues left? Everybody is going to go for EUV ride during the 7nm shift for cost reasons anyway.

Re:EUV

[sexconker]

But EUV is much, much slower. For volume production, you end up with fewer wafers per day.

Re:EUV

[Anonymous Coward]

EUV has been "planned" since mid 2000s and still hasn't appeared; it will happen when it becomes economically feasible to, and the technology is there. Probably first thing we'll see are via layers in EUV, not metal or fin layers; this is because mask defect density is too high (but via layers are sparse, so it may be okay).

Re:EUV

[CanadianMacFan]

But 'm sure that the shift to EUV is expensive which is what it is saying.

Re:EUV

[edxwelch]

Immersion steppers have higher throughput, but as they require more production steps they are effectively slower.

Re:THIS JUST IN:London Westminster Attack Hoax Bus

[Anonymous Coward]

What a loser. Youtube is not how you get sources for outlandish claims. Get a life, please.

EUV

[edxwelch]

"but its approach requires an expensive shift that chipmakers have put off for years: the use of extreme ultraviolet light"

Actually, EUV has been planned for 5nm all along (even for 7nm). It make the process cheaper, not more expensive (by reducing the number of masks)

Re:THIS JUST IN:London Westminster Attack Hoax Bus

[Anonymous Coward]

What a dumb fuck. The videos were taken on site at London.

The video being on youtube facebook or vimeo doesn't change the fact that the attack was yet another hoax.

jimstone.is has been busting fake "attacks" for years.

Re: Enough already!

[Anonymous Coward]

+1 Funny

Did i miss something?

[Anonymous Coward]

Forgive me if this is dumb but what what tech "using today's chips" can I buy that uses 10nm? I thought 14nm was today's commercial tech?

Re: Did i miss something?

[Anonymous Coward]

Samsung has 10nm chips. Exynos 9 8895 is one example.

Re:Did i miss something?

[Anonymous Coward]

http://www.pcworld.com/article/3173376/mobile/samsung-starts-production-of-new-10-nm-exynos-9-series-chip.html

Re:THIS JUST IN:London Westminster Attack Hoax Bus

[MightyYar]

Christ, stop spamming us with his mental illness.

Old technology is old.

[Gravis Zero]

The way we are building processors (primarily lithography) is old, tired and reaching it's limits. What we need now is to focus on figuring out how to make small machines that can work cooperatively to build new things, including copies of themselves. It's been known for quite some time that future of computing is in massive distributed systems, so why aren't we applying the same concepts to manufacturing processors?

Re:Old technology is old.

[fahrbot-bot]

What we need now is to focus on figuring out how to make small machines that can work cooperatively to build new things, including copies of themselves.

Surely, nothing could go wrong with that?

Re:Old technology is old.

[mentil]

We'd probably have to use lithography to create these nanomachines in the first place. Using nano-tweezers to manipulate one atom at a time to create an array of nanoconstructors is infeasible, so it's more likely we'll start with larger machines that can make smaller machines. The problem is they'll be so small they can't rely on optics, yet will require some autonomy; humans can't just dump a sack of copper atoms into a hopper, they need to corral the atoms themselves. Is Intel really going to invent a Universal Constructor and then use it to make smaller computer chips? There are thousands of things they could do with it that would be far more profitable than faster computer chips.

Not exactly news, but..

[SlowCoder]

EUV has been worked on for a while, but at the scales they're ending up on, there's more problems that they don't address in the article. Electromigration starts to become a real problem when the transistors are only a few 10s of atoms wide for example.

Re:Not exactly news, but..

[iotaborg]

EM in Cu is a problem, but who's to say future chips are going to still use Cu?

Re:THIS JUST IN:London Westminster Attack Hoax Bus

[Anonymous Coward]

Stop watching fake news.

Re:THIS JUST IN:London Westminster Attack Hoax Bus

[MightyYar]

I'm trying but you keep posting it.

5 nm welcomes the gig economy

[epine]

It make the process cheaper, not more expensive (by reducing the number of masks)

You are so dating yourself. Just ten seconds on Wikipedia concerning EUV lithography would shave a decade off your musty knowledge.

The whole point of the post-2009 lithographic era is that nothing traditionally used as a benchmark of progress comes for free.

Advantage: fewer masks
Disadvantage: vastly longer step time

Moore's law is still hobbling along, but it definitely lost a testicle circa 2004–2009. You can see it in any honest graph.

Even this article lies a bit.

The chips are down for Moore's law — 9 February 2016

Transistors per chip still kinda going up on the same trend, as always.

But what they don't tell you is that there's an entire CPU inside your CPU devoted to turning those nice transistors off if they work too hard.

Welcome to union rules.

It's just like the gig economy where you only get twenty hours per week on average, but they won't tell which hours ahead of time, so you can't actually get yourself another 20-hour gig to achieve full-time gainful employment.

Official unemployment down; hours and hours of Counter-Strike alone at home (waiting for the phone to ring) way way up.

Re:THIS JUST IN:London Westminster Attack Hoax Bus

[Anonymous Coward]

Are you APK?

Your post has all the classic markings of his shit.

If so please seek help for your moose wang addiction.

If not, still seek help for your moose wang addiction.

All hail Slashdot! ....

[kfh227]

... for suing the correct definition of Moore's law!

3D or miniaturization?

[manu0601]

I understood that when founders anouce improved density nowadays, it is more because of 3D stacking rather than miniaturization.

But the mention of ultraviolet light suggest this time it is indeed about miniaturization. Anyone has more information?