TSMC: 2nm chips by 2025, 3nm this year
Click here to post a comment for TSMC: 2nm chips by 2025, 3nm this year on our message forum
Ryu5uzaku
Intel gets left behind even more... Hnngh it's no good for the general population because TSMC has limited capacity
cryohellinc
I'll just wait for -1nm processors. Now those will be neat.
fantaskarsef
This actually makes sense, if you think about it this way: To get one square nm of chip for a reasonable price, you'll have to go negative, so back through time to when things were a bit different...
KissSh0t
What actually happens when it gets smaller than 1? does it then go something like 0.5?
cucaulay malkin
it's all a matter of units
1000 picometers in one nanometer
0.5nm is more convenient than 500 picometers though
for intel 0.5nm will be "5A"
https://338855-1043895-2-raikfcquaxqncofqfm.stackpathdns.com/wp-content/uploads/2022/04/110816_screenshot-2022-04-11-at-15-51-14-powerpoint-presentation-d1x_mod_3_roundtable_embargoed_april_11_10am_ptpdf.jpg
cucaulay malkin
intel has already secured a chunk of tsmc's 3nm
https://www.techspot.com/news/90780-intel-has-reportedly-secured-majority-tsmc-3nm-production.html
intel has something going on too I'm sure
it takes them longer,their 10nm debuted several years after tsmc's 7nm,but in the end the products are good and that's what matters
https://www.hardwaretimes.com/intel-promises-to-regain-process-leadership-from-amd-tsmc-by-2024-with-1-8nm-18a-process/
according to this,their 2nm will come a year earlier than tsmc's 2nm
SniperX
Can't wait for them to finally break away from the "nanometre" MARKETING naming convention.
As can be seen, many are still confused by the number 🙄
nevcairiel
Intel 4 is also scheduled for 22H2 production (products in early 23), and Intel 3 for 23H2, which might be feasible as Intel 3 is only a tweaked version of Intel 4
Hopefully Intel 4 will also be free from all the issues that plagued Intel 7, since its using EUV, which makes it a bit simpler then the super complex processes required to get Intel 7 working on DUV.
Dragam1337
Allegedly - let's see if it doesn't get delayed by many years aswell... if i were a betting man, i'd put my money on that.
Kaarme
TSMC is a Taiwanese company, so it makes sense for them use nm, an SI unit. Intel, however, is an American company, so it was strange they used nm for as long as they did. It couldn't last forever, naturally. In the true American spirit, they invented a new freedom unit for it. Since Intel is a humble and cordial company, not at all arrogant, they named the new unit... after themselves. Now we have 7 intels, 4 intels, 3 intels...
Denial
Presumably they would just measure in picometers. It's not like it matters though, these are just marketing numbers, nothing about the process actually matches the NM they are suggesting. Which is probably why Intel just dropped it altogether. It's a meaningless name.
Venix
As @Denial said the nm on the naming does not represent the real nm . Pretty much when tsmc 16nm came about and glofos 14nm are more like 20nm finfet in reality they just both named their process like that to seem like they caught up with Intel and marketing the 10nm from Intel are more dense than the "7nm" from tsmc not by much though. If I had a vote in it I would vote for a naming that is the count per square mm ! So like 400k process then the 600k process comes and you instantly know this has 50 % more transistors on the same surface area!
tunejunky
everyone has limited capacity.
but TSMC has far more capacity than you may think. there are brand new fabs for the most advanced processes in Arizona and Taiwan, brand new custom silicon fabs for 3d cache, and more capacity @7n & 5n - in fact far more capacity @7n than Intel because TSMC does more contract work than Intel.
the growth and well being of TSMC is Taiwanese national security and the gov't treats them like that
schmidtbag
As an American, I begrudgingly Like your post.
To be fair though, Intel did this because the use of nm became so twisted that it started to lose meaning.
Ryu5uzaku
Yes and everyone and their mother are using tsmc. While they have capacity it's not enough atm.
For others were talking about Intel's roadmap well they have a lot to prove to be able to keep up or even leapfrog. It's after all their own issues that made 10nm/Intel 7 to be so late.
If Samsung would do better we would have options but the fact that even Intel buys tsmc is sad.
tunejunky
umm excuse me?
you are quite wrong in your assumption about capacity, which has everything to do with yield. TSMC has the highest yields of any silicon fabricator. the difference in production @ node is profound between TSMC and Intel.
furthermore, AMD's uArch leverages that yield with "chiplets" - which are higher yield yet.
in the meantime the factories producing Raphael have spent the last two years pumping out iPhones.
there are more iPhones sold each year than CPU's (by far) and there' never been a shortage of iPhones.
if it was Intel we were talking about i'd be more worried because of their monolithic design (which equals lower yield).
but even there TSMC has excess capacity.
Silva
Up till the early 2000 the number had some kind of meaning, saying the process is "x nm" or is named "x intel" is meaningless now.
As already said, current Intel 10 nm is comparable to TSMC 7nm so intel renamed it to "Intel 7" to look better.
The true comparison is performance per watt, nothing else matters atm.
schmidtbag
Even that isn't so clear, because performance-per-watt is relative. For one thing, PPW can vary simply because of how hot the chip is. Some architectures have better PPW when certain instructions are used. Others have better PPW when comparing mobile versions but not desktop versions (or vise versa). Others are better under very specific workloads, but maybe not under anything else. Some can have better PPW if you just undervolt them. Some get astonishingly bad PPW just by a 5% overclock. The stats can also be skewed because of other factors, like what kind of memory you're using, or the version of the CPU scheduler.
Don't get me wrong, I care a lot about PPW, as well as performance-per-thread and performance-per-clock. As someone who doesn't care about getting into pissing contests or bragging rights, I want the most efficient and well-built product for my needs. But in terms of measuring overall performance at a glance, the number of transistors per square mm seems to be the best general approach.
Venix
+1 although that said you can not even do perf/watt comparison between different architectures (***), you will have the same product produced on both fabs and optimize the best you can for each process. Where there some GPUs that where made on glofo and tsmc on their 14/16 process at some point ?
***: You can totally thought compare 2 different architectures + the process node and end up in a solid conclusion witch combination is more efficient.
cucaulay malkin
transistors per mm2 is what matters afaik
https://twitter.com/iancutress/status/1375053390661705729?lang=en
