TSMC Could Rollout 2nm by 2023

[Vishera]

1 hour ago, ShrimpBrime said:

Then I was hoping maybe Zen 4 (i thought it'd be 5000 series hence filename) would have gone to, say, 88 cores on 5nm...

Sorry to disappoint you but this layout will have very high latency and the others you made.

You want the path to the IO and between dies to be as short as possible so chips that are far from the controller will have latency issues which i expect to be the bottleneck of this layout.

 

[PianoPlayer88Key]

4 minutes ago, Vishera said:

Sorry to disappoint you but this layout will have very high latency and the others you made.

You want the path to the IO and between dies to be as short as possible so chips that are far from the controller will have latency issues which i expect to be the bottleneck of this layout.

 

I was actually the one who made that pic (somehow the quote name tags must have glitched), but maybe that could be somewhat remedied with stacking dice?  (Maybe that would also allow RTX 3090-quantity core counts on Socket AM5, assuming it's the same physical size as and compatible with cooolers from AM4, at like 2nm or 1.4nm? )

 

Let the CPU vs GPU core wars begin! (Come on, when I look for "torch" in emoji database, I want a stick with fire, not a "flashlight".)

 

[Qub3d]

I've liked the use of the term "2nm equivalent". One way I've heard this explained to me is that since the advent of finFET and gate-around-source + atomic layer deposition, cutting edge fab development has essentially become 3D printing transistors out of atoms.

 

Thus it's more effective to think of each process size as being named after the equivalent size of 2D process node would have to be to put the equivalent number of transistors in a given unit squared.

 

this means that these process nodes are almost certainly much larger than seven nanometers on average, but we can just use height much more effectively than previously.

 

In fact I can't find it now but I believe it was a veritasium video that had a physicist straight up state it may be theoretically impossible to build a 2D transistor below about 7 nanometers in size due to quantum effects.

 

-----

 

now I'm of the opinion that quantum computing is not going to magically replace classical architecture anytime soon, if ever. that said our current understanding of how to build transistors may very well change and become more "quantum", in that they rely on constructive and destructive interference to produce effectively the same thing as what a transistor would.

[Vishera]

14 minutes ago, PianoPlayer88Key said:

but maybe that could be somewhat remedied with stacking dice?

Die stacking is limited but a much better solution.

Here is a better layout (I know that some of the traces are disconnected,i was too lazy to fix that )

[tim0901]

2 hours ago, Brooksie359 said:

I was under the impression that smaller transistors result in lower power consumption so even if two processes have similar density the one with the smaller transistors would be more power efficient. 

Density and transistor size are pretty much two ways of measuring the same thing. If you have smaller transistors, you can fit more into the same area and so your transistor density is higher. This is why nodes of a similar density will perform similarly - to achieve the same density their transistors will be pretty much the same size. There may be some variance here, but that will also be dependent on other parameters of node design, such as the design of the individual transistors.

[Mark Kaine]

14 hours ago, Letgomyleghoe said:

you can just keep shrinking forever

By name yes, but technically no. 

 

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

 

And that's one of the main reasons why that is not possible. 

 

 

 

[Doobeedoo]

Fun times, scaling down further with silicon, can't wait to see the final limits, probably another decade potentially. 

[Silentprototipe]

Were close to the end game here. We cant keep shrinking further at some point so I wonder what is gonna be next. Either were gonna see architecture changes instead of transistor changes or just a move to a whole different computing method 

[AndreiArgeanu]

15 hours ago, poochyena said:

I thought we'd need new materials to get much smaller than current processors.

I believe the limit of silicon is actually around 1nm or so, the size of 1 silicon atom is about 0.2nm. So if 2nm is actually 2nm, then we're getting close to the limits of silicon.

[SlimyPython]

5 hours ago, AndreiArgeanu said:

I believe the limit of silicon is actually around 1nm or so, the size of 1 silicon atom is about 0.2nm. So if 2nm is actually 2nm, then we're getting close to the limits of silicon.

Split the atom

Spoiler

boom?

 

[n0stalghia]

9 hours ago, Silentprototipe said:

Were close to the end game here. We cant keep shrinking further at some point so I wonder what is gonna be next. Either were gonna see architecture changes instead of transistor changes or just a move to a whole different computing method 

Efficiency/architecture changes to squeeze way more performance out of existing limitations - I guess the Apple M1 is maybe a forerunner of this. If ARM and good optimization can make such improvements, then this is another solid decade of performance gains.

 

After that... yeah, I have no idea either.

[tim0901]

11 hours ago, Mark Kaine said:

By name yes, but technically no. 

 

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

 

And that's one of the main reasons why that is not possible.

Eh... ish?

 

Heisenberg's uncertainty principle asserts that there's a fundamental limit to how accurately we can measure two linked quantities - in our case that would be the position and velocity - of a given particle. That's all it says. And that fundamental limit is tiny - the uncertainty of the two measurements, when multiplied together, must be larger than 5x10^-35. For reference, 1nm is 1x10^-9. The uncertainty of pretty much everything that humanity has ever created has been many orders of magnitude higher than that limit. Sure in theory we could construct a transistor so small that the accuracies required to do so would be butting up against that limit, but it would never actually work as a transistor. There are other quantum effects that will come into play before then.

 

Generally speaking Quantum Tunnelling is considered the lower limit on how small transistors can be made. The uncertainty principle can be used to describe this mechanism, but it can also be explained more generally by using the properties of the wavefunction itself - how quantum states (and by extension particles) are described in quantum physics. Wavefunctions describe quantum states as probability distributions - the particle it describes is not in a determinate place, rather we have a probability distribution saying where it is likely to be. You can think of it as a bell curve. This curve can overlap an obstruction, and can continue through the other side of it if the obstruction is narrow enough. If it does so, then there exists a probabilty that the particle can be found on either side of the barrier - aka the particle can pass through the barrier.

 

So if that barrier is a transistor gate and the particle an electron, then the electron can skip past the gate if it's narrow enough, causing current leakage (bad). Too much of this and you don't have a transistor anymore, as the difference between an open and closed transistor is too small. This occurs in transistors today, but isn't a huge deal. By the time we get to transistors with features of ~4nm though, we basically reach a limit. (Feature size != node name!! The name of the node is far smaller than any feature of the transistors it uses!)

 

Quote

 

Were close to the end game here. We cant keep shrinking further at some point so I wonder what is gonna be next. Either were gonna see architecture changes instead of transistor changes or just a move to a whole different computing method 

Not necessarily. One idea is to use Tunnel Field Effect Transistors (TFET), which use quantum tunneling, rather than heat, to open and close the transistor gates. This could potentially reduce the operating voltage to ~0.1V and reduce power consumption by ~100x. Unfortunately they're still in the design phase, but progress is being made.

[StDragon]

2 minutes ago, tim0901 said:

Not necessarily. One idea is to use Tunnel Field Effect Transistors (TFET), which use quantum tunneling, rather than heat, to open and close the transistor gates. This could potentially reduce the operating voltage to ~0.1V and reduce power consumption by ~100x. Unfortunately they're still in the design phase, but progress is being made.

 Ok, but how cold will that theoretical chip have to operate at? -200 C or something? If so, that's going to annihilate your power savings. 

[spartaman64]

On 11/16/2020 at 5:41 PM, Letgomyleghoe said:

*cough cough* ARM *cough cough* 

i mean arm isnt a competitor afaik they dont have their own fabs. they are more of a customer

[Mark Kaine]

10 minutes ago, tim0901 said:

Eh... ish?

Tldr: it is a limiting factor, and just because the theoretical limit is smaller, doesn't mean we can reach it with conventional (less than ideal and if we're honest, outdated) tech. 

 

Is what you wanted to say, I believe?

 

 

[Letgomyleghoe]

2 minutes ago, spartaman64 said:

i mean arm isnt a competitor afaik they dont have their own fabs. they are more of a customer

i want saying they are a competitor, just bringing up the point that ARM could enter the consumer space at some point.

 

and AMD doesn't have their own fabs either. 

[tim0901]

3 minutes ago, StDragon said:

 Ok, but how cold will that theoretical chip have to operate at? -200 C or something? If so, that's going to annihilate your power savings. 

Unlikely, they aren't taking advantage of suprtconductivity or anything. In a paper I've found they're testing their device between 223 and 323K (-50C - +50C) which suggests it should work at room temperature. But as I mentioned these are still very much in the research and development stage so a lot could still change.

[tim0901]

3 minutes ago, Mark Kaine said:

Tldr: it is a limiting factor, and just because the theoretical limit is smaller, doesn't mean we can reach it with conventional (less than ideal and if we're honest, outdated) tech. 

 

Is what you wanted to say, I believe?

I believe what i meant to say is that everyone should spend way more money on science so we can figure out alternatives so we never have to reach this limit.

 

No I am totally not biased in this opinion at all...

[spartaman64]

20 hours ago, Vishera said:

Die stacking is limited but a much better solution.

Here is a better layout (I know that some of the traces are disconnected,i was too lazy to fix that )

 

the issue i can see with that is how are you going to cool that? essentially the dies on the bottom are not making contact with the IHS and cooler at all

maybe you can have some weird cpu layout where there are cores on both sides and you just need to get two cpu coolers LUL

[Vishera]

Just now, spartaman64 said:

the issue i can see with that is how are you going to cool that? essentially the dies on the bottom are not making contact with the IHS and cooler at all

maybe you can have some weird cpu layout where there are cores on both sides and you just need to get two cpu coolers LUL

I used a Threadripper package for it so the CPU should have no problems with a TRX40 cooler.

[tim0901]

1 minute ago, Vishera said:

I used a Threadripper package for it so the CPU should have no problems with a TRX40 cooler.

Not the point. If you stack multiple sets of CPU cores on top of each other, then the cores at the bottom aren't touching the IHS. Instead they're covered by another layer of blisteringly hot cores - they have nowhere for the heat they're generating to go.

 

Die stacking so far (eg foveros) has lower powered components such as PCIE and usb controllers at the bottom, with the hot CPU cores on the top to maximise their ability to be cooled effectively. You can't just go stacking CPU cores on top of another without releasing the magic blue smoke.

[Vishera]

1 hour ago, tim0901 said:

Not the point. If you stack multiple sets of CPU cores on top of each other, then the cores at the bottom aren't touching the IHS. Instead they're covered by another layer of blisteringly hot cores - they have nowhere for the heat they're generating to go.

 

Die stacking so far (eg foveros) has lower powered components such as PCIE and usb controllers at the bottom, with the hot CPU cores on the top to maximise their ability to be cooled effectively. You can't just go stacking CPU cores on top of another without releasing the magic blue smoke.

You are right,I didn't think about that.

So instead of that AMD will need higher core density inside the CCDs.

[CarlBar]

3 hours ago, Vishera said:

So instead of that AMD will need higher core density inside the CCDs.

 

More core density means more heat. basically you've got an upper limit on how many watts of power somthing with a given surface area can consume, the thicker it is, (i.e how high it sticks up off the substrate), the lower this is.

[straight_stewie]

On 11/16/2020 at 5:32 PM, Random_Person1234 said:

It looks like we're approaching picometers. So much for 7nm/5nm being the limit.

TSMC's own people say that the node size doesn't actually have anything to do with the node size. In fact, the node size at that time will be at the highest spread against the claimed size since Moore first idealized IC growth patterns.

Manufacturers are using node size as a marketing term, with a very loose connection to transistor density. However, node sizes keep trucking down while actual transistor sizes are starting to plateau.

There are a plethora of links to reputable tech news organizations that also say this. Here are a few:

• https://www.pcgamer.com/chipmaking-process-node-naming-lmc-paper/
• https://hexus.net/tech/news/cpu/145645-intel-14nm-amdtsmc-7nm-transistors-micro-compared/
• https://www.hpcwire.com/2020/06/01/10nm-7nm-5nm-should-the-chip-nanometer-metric-be-replaced/

 

 

[Mihle]

6 hours ago, Vishera said:

You are right,I didn't think about that.

So instead of that AMD will need higher core density inside the CCDs.

One thing AMD possibly could do eventually is to place come core chips on top of the IO chip.

 

I did read somewhere that someone suggested the possibility to have L3 cache (or L4) in the IO chip, having the IO die be larger than it is now, and stack the core chips on top of it.

 

Personally I don't have knowledge to know of that is possible or not. Would doing that make the cache in IO chip be too slow to be L3 cache or not?