• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

TSMC will start mass production of 3nm chips in the second half of 2022

tusharngf

Member





Taiwanese chipmaker, TSMC, will commence the mass production of 3nm chips in the second half of 2022. This is the latest report with respect to the development of TSMC’s 3nm manufacturing process. In addition, the monthly production capacity of this process will be 55,000 pieces. According to TSMC’s Chairman, Liu Deyin, the number of employees in the Tainan Science Park will reach about 20,000 when the 3nm chip starts mass production. The current number is 15,000.

In August of this year, Mi Yujie, senior vice president of TSMC, said that the company has plans to continue to provide meaningful node improvements until N3 and below. TSMC expects N3 to have the latest and most advanced node in 2022. Compared with N5, the benefits are similarly small, performance is only improved by 1.1-1.15 times, and power consumption is reduced by 1.25-1.3 times. These gains are relative to N5, not N5P.

If you compare the performance with the 7nm process, N3 performance improves by 1.25 times – 1.35 times. Also, the power consumption reduces by 1.55 times-1.6 times under the same usage. In these comparisons, all the multiples you see assume an idealized transistor, which does not necessarily match the actual products manufactured by AMD, or Intel.

TSMC N3 will continue to use FinFET fin field-effect transistors instead of transitioning to GAA wraparound structure field-effect transistors. This is different from Samsung, which has already stated that it will use GAA at the 3nm node.

SAMSUNG TO MASS-PRODUCE 3NM CHIPS IN 2022: STRIVING TO OVERTAKE TSMC
Samsung and TSMC are the only two companies that have the capacity to make chips with the 5nm process. However, Samsung seems to be in the shadow of TSMC but it is putting up a good fight. According to recent reports, Samsung Electronics is struggling to catch up with TSMC. The South Korean manufacturing giant plans to mass-produce 3nm chips in 2022.


Source: TSMC will start mass production of 3nm chips in the second half of 2022 - (gizchina.com)
 
I don't follow the semiconductor industry super closely, but has there been any talk lately about using alternative materials like germanium?

I found this IEEE Spectrum article from the end of 2016.
I'm sure it's been considered. I'd say if we can't get much smaller without a complete paradigm shift they are going to HAVE to find a way to make transistors more efficient. I'd wager germanium hasn't been used because it probably wasn't as cost efficient in the past. Efficiency at 3nm will probably be the only way to move forward in the next few years as we are still a ways away from Quantum Computing. Even then, we are just hitting limits on how many transistors can physically fit on silicon so some creative solutions are gonna have to come forward to make any progress.
 

Jagz

Member
It's going to be exciting seeing these used in small Switch-competitor UMPCs, like the GPD WIN series, etc. Maybe Samsung will release their own UMPC that competes against the Switch, since they do make laptops, after all.
 

LordOfChaos

Member
Good. It's gonna be interesting in 2 years or so as we Moore's law is finally pretty much at it's end and manufactures are gonna have to find different ways for efficiency. It's gonna be an interesting time to be an engineer..

Sorta. Fab process names are so decoupled from what's actually going on inside the chip that we still have a ways to go. When they say like, shrinks will be harder past 3nm transistors, they do mean 3nm transistors, not what TSMC/Samsung etc decided to name a 3nm process because 3nm is the minimum feature size on some area of the chip. Transistors are three dimensional objects and have some ways to go, just look at fin lengths.

I'm with this chip god

 
Sorta. Fab process names are so decoupled from what's actually going on inside the chip that we still have a ways to go. When they say like, shrinks will be harder past 3nm transistors, they do mean 3nm transistors, not what TSMC/Samsung etc decided to name a 3nm process because 3nm is the minimum feature size on some area of the chip. Transistors are three dimensional objects and have some ways to go, just look at fin lengths.

I'm with this chip god

If anyone would know, it'd be Lex Friedman. So I'd defer to anything him(and Jim Keller) would say as they would absolutely know more than I would. It's crazy to me that we have to take the size of atom's in consideration to design now. When I learned about transistors/resistors/etc. years ago in the military, we never talked about it so it's super interesting to learn about.
 
Last edited:

diffusionx

Member
Pretty crazy. Given how Intel has pushed 10 NM and 14 NM I wonder how far they (and and) will be able push these.

From what I understand, Intel’s 14++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++nm and the 7 or 5 or whatever that TSMC talks about are comparable in a lot of ways. These chip densities have become marketing speak like everything else unfortunately. Intel actually isn’t 5 years behind TSMC or Samsung.
 
Last edited:
Good. It's gonna be interesting in 2 years or so as we Moore's law is finally pretty much at it's end and manufactures are gonna have to find different ways for efficiency. It's gonna be an interesting time to be an engineer..
Multi-chip modules and 3D stacking just to name a few. Looking forward to what Nvidia brings to the table with Hopper.
 

Trimesh

Banned
From what I understand, Intel’s 14++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++nm and the 7 or 5 or whatever that TSMC talks about are comparable in a lot of ways. These chip densities have become marketing speak like everything else unfortunately. Intel actually isn’t 5 years behind TSMC or Samsung.

TSMC's original "7nm" process had slightly lower transistor density than Intel's "10nm", although their later N7+ process was slightly higher. TSMC also demonstrated an ability to manufacture these parts in volume with presumably decent yields - which Intel's reluctance to mainstream their 10nm process suggests they didn't. It would be very interesting to know what exactly went wrong, but Intel are obviously not talking.
 
I hope PS5 Pro and XsX Pro are going to use 3nm (or 3nm+) instead of 5nm. They can release the console Holiday 2024.

They can have an alpha GDK kit for PS5Pro and XsXPro using 5nm and give to developers holiday 2022 to give them time to develop next next gen games.

Then PS6 and XsX^2 are going to use 1nm.
 
Maybe we can get PS5 Pro in 2023 if it ramps up well.

There won't be a PS5 Pro.

3nm will be way too expensive for whatever timeframe a PS5 Pro launches (2023 or 2024), and if they double the GPU size then the "fast & narrow* design philosophy gets thrown out of the window because that's twice as many CUs, SEs and SAs to parallelize work distribution through. That is of course if you 100% buy the talking points into that type of design philosophy.

Not to mention, a 72 CU chiplet doubles the silicon price, in addition to the increase in price 3nm brings over the 7nm process. Then you have to consider what memory will be available to feed it with data, what capacities, the prices, even when also factoring in things like Infinity Cache. Lastly, if you also want to double drive storage space AND increase performance thereof, you're looking at a system easily 2x the price of PS5.

There's a much stronger chance a mid-gen refresh will be along the lines of a Slim-style model on the N6 process, which keeps the same performance but with lower power consumption and a density shrink (I believe). It should also be cheaper than 7nm; that would give room for a "premium" SKU packing in expanded amount of storage.

I think any other mid-gen refreshes will be: smaller-model Series S with less power consumption and possibly 2x storage, some portable/tablet style Series S-based device, a PSVR2 with some kind of ability to play limited VR games without a PS5 (but functioning as an enhanced VR device with PS5), and a cost-reduced Series X on N6, in two versions: one without a disc drive and 1 TB storage, and a pricier SKU with a disc drive and 2 TB storage.

Probably don't expect the average price for any lower-end SKU replacements to base units to be lower by more than $100, either. All the other stuff like 3nm/3nm EUV, fully redesigned GPUs, massively bigger-bandwidth storage solutions etc., don't expect that until 10th-gen systems hit.

I hope PS5 Pro and XsX Pro are going to use 3nm (or 3nm+) instead of 5nm. They can release the console Holiday 2024.

They can have an alpha GDK kit for PS5Pro and XsXPro using 5nm and give to developers holiday 2022 to give them time to develop next next gen games.

Then PS6 and XsX^2 are going to use 1nm.

Realistically it's impossible. Die shrinks are getting more expensive, not less, and in the end probably not worth the costs. PS5 and Series X aren't really hurting for computational power whatsoever, it's storage space that's their biggest enemy ATM. So the best way to tackle that with a mid-gen refresh is to go to 6nm process, get the power consumption savings there while keeping same performance envelopes, and put a lot of the budget towards increased storage space instead.

So I'm thinking mid-gen refreshes will be more along the line of Slim models versus anything major in pushing higher performance, outside of maybe the SSD I/O on the systems getting a slight revision to eke out a tad more performance there. There's a much better chance we'll see some more "interesting" mid-gen refreshes in the form of PSVR2, and possibly a portable/tablet-styled Series S device.
 
Last edited:
Thankfully we are at point where developers have enough to power to create pretty much anything they want with excellent visual quality. Hardware and visuals will keep improving but I don't think there is any great rush with the diminishing returns we are seeing.
 

Elios83

Member
This company is on fire with their advanced manufacturing processes.
It's definetly good news for for AMD and Sony/MS.
 
nothingstopsthistrain.gif

From there on I think it will be all about MCM design and 3D stacking. Snd increased prices, definitely increased prices.



What if 2020 is just a prelude to 2021?? :O
2021... If you just swap the last two digits... 2012! It's the end of the world!

Also 3 different numbers in that number... 3... Half Life 3 in 2021 (or 2012 part 2) confirmed.
 

Moochi

Member
I'm betting on a mid-gen refresh with roughly the same graphics and processor, more storage, and the key feature being addition of reworked chip design integrating an AI upscaling/tensor core module.
 
Top Bottom