TSMC confirms 3nm tech for 2022, could enable epic 80 billion transistor GPUs

[tusharngf]

TSMC has confirmed its 3nm production node is on track for full mass production in the second half of 2022, according to Chinese tech site ItHome (in Chinese). TSMC reckons its 3nm node will pack in somewhere north of 250 million transistors per square millimetre of silicon, making it at least two and half times more dense than Intel's latest 10nm node. In theory, TSMC's 3nm tech could enable a GPU three times more complex than AMD's new Radeon RX 6000 Series chips



TSMC, of course, makes all of AMD's high performance Ryzen CPUs and Radeon GPUs. Until recently, it also produced Nvidia's top graphics chips, too. Advances like TSMC 3nm tech matter because they allow for more complex, faster computer chips. Like, you know, CPUs and GPUs.

But just think about it. AMD's Navi 21 GPU, the chip inside the Radeon 6800 and 6900 Series boards, clocks in just under 27 billion transistors using 7nm. So a 3nm GPU of the same size could pack around 80 billion transistors if they really pushed it.



More at
Source: https://www.pcgamer.com/tsmc-confir...could-enable-epic-80-billion-transistor-gpus/

 

[Fat4all]

so like eight gamecubes taped together?

 

[Darrilium]

so like eight gamecubes taped together?

Click to shrink...

Eight thousands.

 

[Fat4all]

Eight thousands.

Click to shrink...

that's a lotta ducktape

 

[JFoul]

Meanwhile at Intel....

 

[ILikeFeet]

that's gonna be an expensive bag of chips

 

[Cow Mengde]

Meanwhile at Intel....

Click to shrink...

14nm ++++++++++++++++++++++++++++++++++

 

[Unkindled]

Meanwhile at Intel....

Click to shrink...

14nm +∞

 

[Madao]

just when i have plans to upgrade my current PC next year. now i don't know about waiting even more or doing a second upgrade in 5-6 years.

 

[Ghostsonplanets]

While these numbers are great, let's not forget that not everything in chips are transistors. Logic takes a huge chunk of that and doesn't scale the same as transistors. It isn't simple as this "3nm GPU = 80 billion transistors". It will be quite less than that.

 

[Akabeko]

just when i have plans to upgrade my current PC next year. now i don't know about waiting even more or doing a second upgrade in 5-6 years.

Click to shrink...

The first year of production will most likely be almost entirely gobbled up by Apple so I wouldn't let this news dissuade you from upgrading.

 

[neoak]

TSMC numbers are marketing, so this would be, based on history, equivalent to a theoretical Intel 5nm process.

Leakage is gonna be a pain at this sizes

 

[Deleted member 2533]

just when i have plans to upgrade my current PC next year. now i don't know about waiting even more or doing a second upgrade in 5-6 years.

Click to shrink...

I think now's a really good time to upgrade. DDR5 is still a little ways off, and first gen cpu/mobos probably won't be mature right off the bat; raytracing has been around for a while, so the 3000 series GPUs give good support.

Basically, anything on the horizon will probably mean a completely different set of architectures that will be overpriced and underperforming.

 

[acebeam]

This is interesting :

 

[acebeam]

 

[nogoodnamesleft]

what happens when we cant get any smaller? serious question

 

[ekim]

what happens when we cant get any smaller? serious question

Click to shrink...

IIRC 3nm are not really 3nm. There is more to come.

 

[nogoodnamesleft]

IIRC 3nm are not really 3nm. There is more to come.

Click to shrink...

oh no no i mean what happens after 1nm? and then 0.5nm? or whatever? what happens when we hit limits? is this actually something that will happen?

 

[ekim]

oh no no i mean what happens after 1nm? and then 0.5nm? or whatever? what happens when we hit limits? is this actually something that will happen?

Click to shrink...

This is going to take a while and until then we might be able to really use quantum computing.

 

[RCSI]

what happens when we cant get any smaller? serious question

Click to shrink...

Chips will still be made, some progress may be made in other areas, but the jump in performance wont be as large. For example, Intel 14nm processors may be an indicator of what that may look like.

Edit: As Ekim indicated. I also assume this is decades off and that there is still scaling on the roadmaps. Another point, 3nm, 2nm etc is not the size of the components on the chip, that's marketing stuff.

 

[Bunzy]

oh no no i mean what happens after 1nm? and then 0.5nm? or whatever? what happens when we hit limits? is this actually something that will happen?

Click to shrink...

we move away from silicon, or some sort of stacking

 

[19thCenturyFox]

oh no no i mean what happens after 1nm? and then 0.5nm? or whatever? what happens when we hit limits? is this actually something that will happen?

Click to shrink...

I mean you can still have performance gains of 30% or higher on the same process node. My guess is the longer the nodes stagnate the more performance they'll keep squeezing out of the tech with architectural improvements. For a few years it won't make much of a difference and from there on out it's going to be all about proprietary tech and feature sets.

 

[elenarie]

what happens when we cant get any smaller? serious question

Click to shrink...

It will mean that we would have to figure out a new way of making chips and transistors. 3nm are already pushing the physical world and laws limts of using extreme ultraviolet lithography for the fabrication process.

 

[nogoodnamesleft]

thanks for replies guys! i find this all fascinating. reaching limits etc. when that happens will that be like a once in a lifetime kind of breakthrough? ie "quantum computing"?

 

[gozu]

oh no no i mean what happens after 1nm? and then 0.5nm? or whatever? what happens when we hit limits? is this actually something that will happen?

Click to shrink...

Lol. Honest answer? We're pretty fucked right now. Diamond or other carbon replacements for silicon never panned out so we've been doping silicon and stretching things as much as we can, at enormous monetary cost.

We're getting massacred by thermal limits and almost every single foundry has floundered and gone out of business before reaching single digit nm. The only exceptions are TSMC and Samsung. And Samsung lags far behind.

There is talk about 3D chips but the thermals are just terrible for them. Really terrible. We have a hard enough time cooling a "2D" chip now. Spitballing? Cooling a 3D one will require hundreds or thousands of pipes within the chip, using some sort of next-gen fluid-like cooling that is one order of magnitude better than what we have now, to make up for the order of magnitude increase in heat.

We were supposed to have ditched silicon and blasted through the 10 Ghz barrier by 2020. What happened?


Welp, blame yourself or God, I guess, as Delita would say.

 

[Nali]

Nothing makes me feel quite so much like we're living in some incredible sci-fi future like hearing how many millions of transistors we can cram into a square millimeter to produce a common, everyday consumer product. Just unfathomably small scales at extremely high complexity.

 

[ItWasMeantToBe19]

Can these even run faster than 7nm? I though heat was a massive issue smaller than 7nm.

 

[DopeyFish]

Lol. Honest answer? We're pretty fucked right now. Diamond or other carbon replacements for silicon never panned out so we've been doping silicon and stretching things as much as we can, at enormous monetary cost.

Click to shrink...

well we still possibly have graphene to let us down

maybe

 

[gozu]

Can these even run faster than 7nm? I though heat was a massive issue smaller than 7nm.

Click to shrink...

Speed is not the end all be all. You want more transistors on the same surface. They can do work in parallel.

 

[RCSI]

Nothing makes me feel quite so much like we're living in some incredible sci-fi future like hearing how many millions of transistors we can cram into a square millimeter to produce a common, everyday consumer product. Just unfathomably small scales at extremely high complexity.

Click to shrink...

The EUV machines that etch the chips seemed more impressive than the chips themselves. Posted already, but it's still insane 6m - 6:49:

 

[modiz]

This is interesting :

Click to shrink...

of course should be noted that these include binned chips so the actual number of consoles manufactured from these is much lower.
that is still a ton of PS5 chips in a single quarter.

 

[Pachinko]

My personal takeaway from this is that Intel is in some serious trouble , their 10nm update will only catch them up to the competitions 7nm at best and with all the delays it also won't be ready until this 3nm stuff is pretty much landing. They need an engineering miracle and I hope they find it just to continue being competitive.

 

[monketron]

So PS5 / XSX slims in 2023!

 

[caiocmsouza]

TSMC has confirmed its 3nm production node is on track for full mass production in the second half of 2022, according to Chinese tech site ItHome (in Chinese). TSMC reckons its 3nm node will pack in somewhere north of 250 million transistors per square millimetre of silicon, making it at least two and half times more dense than Intel's latest 10nm node. In theory, TSMC's 3nm tech could enable a GPU three times more complex than AMD's new Radeon RX 6000 Series chips



TSMC, of course, makes all of AMD's high performance Ryzen CPUs and Radeon GPUs. Until recently, it also produced Nvidia's top graphics chips, too. Advances like TSMC 3nm tech matter because they allow for more complex, faster computer chips. Like, you know, CPUs and GPUs.

But just think about it. AMD's Navi 21 GPU, the chip inside the Radeon 6800 and 6900 Series boards, clocks in just under 27 billion transistors using 7nm. So a 3nm GPU of the same size could pack around 80 billion transistors if they really pushed it.



More at
Source: https://www.pcgamer.com/tsmc-confir...could-enable-epic-80-billion-transistor-gpus/

Click to shrink...

RDNA 3 and Zen 4 in 2022, hype!!!

 

[Deleted member 20831]

What will be the thermals at that density?

 

[OnionPowder]

Can these even run faster than 7nm? I though heat was a massive issue smaller than 7nm.

Click to shrink...

The main issue from what I could gather was that the chips were too thin to contain logic gates. After a certain size the electrical currents jump past logic gates which give the CPUs an inaccuracy. The new crazy shit they have to build the dies is why they can go under 5nm, which was previously a theoretical limit. They use stuff that's measured in picometers (1/1000 of a nanometer)

 

[Evilmaus]

so like eight gamecubes taped together?

Click to shrink...

Mmm, still not following. What is that in DBZ characters?

 

[Jade1962]

This is interesting :

Click to shrink...

Thanks for the info. Just came across this post in another thread. Do we know if Q4 for AMD is Jan-Mar or Oct-Dec?

 

[1-D_FE]

I mean you can still have performance gains of 30% or higher on the same process node. My guess is the longer the nodes stagnate the more performance they'll keep squeezing out of the tech with architectural improvements. For a few years it won't make much of a difference and from there on out it's going to be all about proprietary tech and feature sets.

Click to shrink...

This is the problem Intel has been stuck with. Jim Keller has some interesting quotes about this. How it's good until it's not... and then you're in deep trouble (current Intel).

On a related note, I had no idea until recently that Jim Keller abruptly resigned from Intel during the summer. The reset thread was like 8 posts long. Was probably posted in the middle of the night and dead and buried by the time I checked.

 

[sifi36]

Thanks for the info. Just came across this post in another thread. Do we know if Q4 for AMD is Jan-Mar or Oct-Dec?

Click to shrink...

For both AMD and TSMC, Q4 would be October to December. Therefore these chips are most likely for January to March shipments. One would think there is some correlation between these numbers and shipments this holiday, maybe even a wider split if Microsoft started manufacturing later.

 

[Prototype Viktor]

time for PS Vita 2

 

[Jade1962]

For both AMD and TSMC, Q4 would be October to December. Therefore these chips are most likely for January to March shipments. One would think there is some correlation between these numbers and shipments this holiday, maybe even a wider split if Microsoft started manufacturing later.

Click to shrink...

Appreciate the response. With starting later I would think MS would then increase it's orders for Q4 to make up for the later start early next year.

 

[Skel1ingt0n]

Eh, I'm all about that 14nm++++++++++++++++++++++++++++++++++ ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

 

[Diarrhea Day-Lewis]

My hope would be for a 3nm PS5 slim in 2023 at $299 or $349 that's almost half or 30% smaller than the launch unit.

I assume 3nm would also enable them to shrink the fan, case, motherboard (with less ram chips too), PSU, heat sink, etc

 

[DekuBleep]

Is iPhone 14 in 2022 gonna be 3nm then?

 

[ILikeFeet]

My hope would be for a 3nm PS5 slim in 2023 at $299 or $349 that's almost half or 30% smaller than the launch unit.

I assume 3nm would also enable them to shrink the fan, case, motherboard (with less ram chips too), PSU, heat sink, etc

Click to shrink...

I honestly question if 3nm can get systems to be so cheap

 

[Ada]

These guys are wizards, soon they'll be making the worlds chips. Pretty much every hot device this Christmas is being bottlenecked by their supply.

 

[Euler007]

well we still possibly have graphene to let us down

Click to shrink...

I remember graphene being a hot topic when I was in engineering school circa 2000. I wonder if I'll be retired by the time it actually impacts any industry. That and wondering if the battery breakthrough happens before or after I retire.

 

[Diarrhea Day-Lewis]

I honestly question if 3nm can get systems to be so cheap

Click to shrink...

I imagine at worst a 3nm APU in 2023 would cost console makers the same as today's 7nm models just because I doubt they agreed to a fixed rate that includes a price hike.

So if it costs Sony $120 per APU today, at worst $120 in 2023 on 3nm. But I also think a smaller case, smaller heatsink, smaller fan, smaller motherboard, less RAM Chips (like say 4 4GB instead of 8x2GB), cheaper SSD prices due to wider adoption can shave a good amount off too. Enough for $300? Idk I assume they will hit $299 pricing if only for digital model because they clearly wanted to hit $399 with digital model.


But who Knows I'm not right on these things often

 

[ILikeFeet]

I imagine at worst a 3nm APU in 2023 would cost console makers the same as today's 7nm models just because I doubt they agreed to a fixed rate that includes a price hike.

So if it costs Sony $120 per APU today, at worst $120 in 2023 on 3nm. But I also think a smaller case, smaller heatsink, smaller fan, smaller motherboard, less RAM Chips (like say 4 4GB instead of 8x2GB), cheaper SSD prices due to wider adoption can shave a good amount off too. Enough for $300? Idk I assume they will hit $299 pricing if only for digital model because they clearly wanted to hit $399 with digital model.


But who Knows I'm not right on these things often

Click to shrink...

wouldn't less chips affect bandwidth?