News Intel looks beyond silicon, outlines breakthroughs in atomically-thin 2D transistors, chip packaging, and interconnects at IEDM 2024

[Admin]

Today, the Intel Foundry Technology Research team announced technology breakthroughs in transistor, interconnects, and packaging technology, among others, which it will unveil in seven of its own papers, along with two more papers in collaboration with industry partners like imec, at the IEEE International Electron Devices Meeting (IEDM) 2024 conference.

Intel looks beyond silicon, outlines breakthroughs in atomically-thin 2D transistors, chip packaging, and interconnects at IEDM 2024 : Read more

 

[usertests]

Intel: A PowerPoint Company

 

[rluker5]

Ruthenium isn't cheap. A quick search gave a price of $575.73 per troy ounce. Although they would probably just use a fraction of a gram per chip, the price would rise if they used a lot.

 

[steve15180]

It's great to announce all of these new ideas, it's a whole different thing to bring these great ideas to productions. Intel 20A anyone?

 

[bit_user]

Yes, instead of a physical insulative barrier around the wire to keep electrons moving to the correct places, air is used as the insulator (air has a dielectric constant of around 1.0), which improves capacitance quite a bit as well (Intel claimed a 17% improvement in capacitance with 14nm).

So, is this only the top layer, or else how is the air pocket protected? Is the resulting die somehow sealed? If not, I'd worry about sensitivity to altitude, humidity, and use in environments with a different gas composition than the atmosphere.

 

[bit_user]

Intel: A PowerPoint Company

They can't do only implementation or fundamental research. There's a "pipeline" analogy, where new ideas need to be discovered, tested, refined, and evaluated for use in production. These things need to happen many years in advance of when you want to actually use them in volume manufacturing.

If you put 100% of your people on just your current node, then the business will fail as demand for that node slumps, because you haven't been investing in what's next and there's no substitute for doing all of that work in advance. Even working only on the current + next nodes isn't enough time to develop and refine all the ideas necessary to make a competitive node. Competing means having people focused on long-range, medium-range, and refining what's new and next.

Also, this stuff doesn't happen in isolation. There are many equipment, materials, and software suppliers in the ecosystem. Collaborating with them is essential. If Intel doesn't win their mindshare and attention, TSMC, Samsung, Rapidus, and other will. That would leave Intel potentially with a new node but no supply chain or supporting toolchains.

It's great to announce all of these new ideas, it's a whole different thing to bring these great ideas to productions. Intel 20A anyone?

I'd love to have more insight into what happened with that. As the history books get written, Intel 4 and 20A might go down as unfortunate detours and ultimately mistakes.

As internal test nodes, maybe they were justified, but I think there's a lot of work that happens when refining a test node for use in volume production that basically turned out to be wasted effort that could've been better spent on Intel 3 and 18A.

 

[thestryker]

Ruthenium isn't cheap. A quick search gave a price of $575.73 per troy ounce. Although they would probably just use a fraction of a gram per chip, the price would rise if they used a lot.

I know pretty much nothing about the most cost effective methods of refinement and whether there's any scale to be had, but it's possible pricing could stabilize if there is scale. Some quick looking found that the costs have already skyrocketed as usage has increased (3-5x increase over the last decade). Seems like HDDs had used it as well so as that market collapsed so did pricing (pre-2013) and it's still not back to the peaks it had then.

I'd love to have more insight into what happened with that. As the history books get written, Intel 4 and 20A might go down as unfortunate detours and ultimately mistakes.

As internal test nodes, maybe they were justified, but I think there's a lot of work that happens when refining a test node for use in volume production that basically turned out to be wasted effort that could've been better spent on Intel 3 and 18A.

I wonder if the confluence of buying more DUV equipment (covid spike), having to buy all of the EUV equipment, High-NA EUV delays and node delay undermined the strategy. The only way I see it making sense is if you nail them all and then shift to the next node for the next product cycle. I hope this is what's behind the 20A cancelation, but until there's volume 18A I'm not hanging my hat on anything.

 

[thestryker]

The advances with design are expected, but it has been really interesting to see how rapidly packaging improvements have become as important. Hopefully some of this will pan out on the scale. manufacturing side as that seems to be the biggest problem with a lot of the fabrication breakthroughs.

Part of me wonders if we would be having this shift right now had 450mm wafers happened.

 

[dehjomz]

This is all well and good. But when is intel going to deliver a chip based on these designs? As well as a chip that beats the 9800x3d in gaming? Right now Intel is being seen as second best in x86 products. So the design and manufacturing teams need to pull a rabbit out of a hat to tackle AMD (and by extension, TSMC).

 

[watzupken]

What is the point if you cannot scale it for sale to make money? Isn’t this like Samsung where they kept marketing GAA and whatever XXnm, but no customer because the yield is too poor. It’s all on paper, but can’t make it to the market.

 

[bit_user]

This is all well and good. But when is intel going to deliver a chip based on these designs?

The packaging improvements seem like they could probably apply to products the soonest, which I'd guess might be as early as 2026 or 2027. The wire/insulator improvements seem probably like an intermediate-range thing that could probably apply to the node after 14A. The planar transistor might be even further off.

As well as a chip that beats the 9800x3d in gaming? Right now Intel is being seen as second best in x86 products.

They claim to be delivering new microcode that should improve gaming performance on Arrow Lake, but I wouldn't count on it beating the 9800X3D. So, that means you'll probably have to wait at least until Nova Lake. However, Bartlett Lake (a Raptor Lake derivative) should launch in Q1 and could turn out to be a better gaming CPU than even Arrow Lake with updated microcode.

So the design and manufacturing teams need to pull a rabbit out of a hat to tackle AMD (and by extension, TSMC).

Intel's next node is called 18A and Nova Lake is supposed to use it. So, they're working on it.

 

[thestryker]

As well as a chip that beats the 9800x3d in gaming?

People don't seem to grasp how unimportant the retail market (where most X3D sells) is to Intel's bottom line. The X3D parts are fantastic, but Intel sells so much more volume in laptops and OEM that the cost of making a competitor has never made sense. Hopefully the cost factor will make sense now that they've shifted to tiles for everything, but there are definitely no guarantees.

 

[dlheliski]

Ruthenium isn't cheap. A quick search gave a price of $575.73 per troy ounce. Although they would probably just use a fraction of a gram per chip, the price would rise if they used a lot.

A wafer is about 900cm^2 including blank space to fill the square, and a layer is about .1u, or 1e-5cm thick. This gives .009cc or about .1 gram of Ru per wafer. $570/oz is about $20 per gram, so maybe $2 per wafer for a layer of Ru on a wafer that will be at least $20K to produce.

 

[dehjomz]

People don't seem to grasp how unimportant the retail market (where most X3D sells) is to Intel's bottom line. The X3D parts are fantastic, but Intel sells so much more volume in laptops and OEM that the cost of making a competitor has never made sense. Hopefully the cost factor will make sense now that they've shifted to tiles for everything, but there are definitely no guarantees.

I’d argue that the Reputation in DIY counts for something.

 

[bit_user]

so maybe $2 per wafer for a layer of Ru on a wafer that will be at least $20K to produce.

Thanks for working that out, but I'd point out that there's surely some waste and prices obviously aren't fixed. If demand significantly increased, supply might have difficulty keeping up. Then again, if it was used in hard disks, then perhaps supply isn't an issue?

 

[stuff and nonesense]

What is the point if you cannot scale it for sale to make money? Isn’t this like Samsung where they kept marketing GAA and whatever XXnm, but no customer because the yield is too poor. It’s all on paper, but can’t make it to the market.

Therein lies the problem for all engineering. Design, prototype, modify, prototype and hopefully serialise production.

If the problems seem insurmountable (was it 10xxxxxx or 14xxxxx) go back to design and start again.

A focus on what needs to be achieved must be maintained, that requires good people in all posts who are invested in progress. Those people all the way up the chain must be honest with the results of their labours and be prepared to admit that a given path may not be achievable in that form.

Let’s hope that the optimism in the PP slides is warranted.

 

[craigss]

So, is this only the top layer, or else how is the air pocket protected? Is the resulting die somehow sealed? If not, I'd worry about sensitivity to altitude, humidity, and use in environments with a different gas composition than the atmosphere.

Intel have invented.........Vapourware! 😱

 

[dlheliski]

Thanks for working that out, but I'd point out that there's surely some waste and prices obviously aren't fixed. If demand significantly increased, supply might have difficulty keeping up. Then again, if it was used in hard disks, then perhaps supply isn't an issue?

OK lets do some more math. Start with .1g per wafer. A very large fab is 30K wafers per month, or 3kg of Ru per month, 36kg per year. World supply of Ru is 30,000kg per year, enough for 800 fabs all using this.