DegustatoR
Legend
TSMC wafer pricing
- Thread starter Silenti
- Start date
DegustatoR
Legend
bigger is always better
A16 is almost certainly just N2+ BSPD. I don't think it is disappointing as such given that the timing of A16 and N2P are similar. It seems to be a divergence in cost models for different applications. From TSMC's press release - "A16 ideal for HPC products with complex signal routes and dense power delivery networks", and given TSMC has chosen the most complex application of BSPD, it would certainly be a significant increase in wafer cost. TSMC has a diverse base of customers and has to cater to all possible applications/products. For those who do not need the absolute benefits of A16, N2P would be a cheaper alternative (and IP compatible with N2). Whereas for HPC, A16's benefits would probably outweigh the additional costs. Given the low number of external customers, Intel would mostly have their own internal CPU/GPU tiles on 18A, which would benefit from BSPD and hence makes sense to have on 18A. Even then I believe it is optional and can be dropped if not required.
TSMC is exploring high NA for A14 or whatever their next node is (which I expect to go to HVM in 2H'27) but they have a much higher volume than Intel and would need a significantly higher number of HNA machines. ASML would need to have the capacity to supply. Perhaps we could see them play it safe and go for High NA for A14P instead.
D
Deleted member 2197
Guest
May 6, 2024
wccftech.com
TSMC's Entire CoWoS Supply Reportedly Reserved By NVIDIA & AMD Until 2025
AMD & NVIDIA have reportedly reserved TSMC's entire CoWoS production for the next two years as both firms aggressively compete in the AI race.
wccftech.com
I was wondering the same and was flipping back and forth between wafers and lots.
They don't specify a time period though, so...
I would have to assume it is per month and a unit is a wafer.
Edit- I was initially thinking lots because I misread the article and thought they were talking about ~50k units per year.
Edit- I guess it helps if you go to the source and ignore the shoddy journalism.
台積電先進封裝產能被訂光 輝達、超微一路包到明年 | 產業熱點 | 產業 | 經濟日報
AI應用百花齊放，兩大AI巨頭輝達（NVIDIA）、超微（AMD）全力衝刺高效能運算（HPC）市場，傳出包下台積電今、明...
Literally from the article:
*Chip density published by TSMC reflects 'mixed' chip density consisting of 50% logic, 30% SRAM, and 20% analog.
Well ok, fair enough if that's the case. It just makes the numbers beyond meaningless.
Doesn't it explain why the marketing numbers are low and getting lower?
They are giving a more realworld estimate of density scaling of a "typical" design.
To my knowledge, analog stopped scaling about two decades ago and SRAM has effectively stopped scaling below 5nm.
DegustatoR
Legend
Not really. 50% of logic is a sizeable chunk in a formula and yet we're still looking at ".1x" density improvements. N5 to N3 is the last transition where there are noticeable improvements, and even that will be dissolved a bit because the difference between the most dense N5 version and the most dense N3 version will probably be smaller than between N5 and N3E from the spreadsheet.
Scaling is dead. There are no more "free transistors".
But you are comparing that to previous numbers that didn't use this more "real world" formula.
TSMC listed N3 vs N5 at 1.7x but according to their "new" numbers N3E vs N5 is only 1.3x
N2's 1.15x = 1.3x
A16's 1.07x-1.1x = 1.14x-1.2x
So that would end up being >1.5x according to their old method.
Edit- They are in their incremental improvement phase since they need something new to offer every ~12-18months like they promised.
Last edited:
The "old formula" actually featured SRAM shrinkage though. Like 10 to 7nm was 0.83x fin pitch and 0.82x gate pitch for 47% higher logic density, while also HD SRAM cell size shrunk from 0.042 um^2 to 0.027. So a mix of 1.47x logic and 1.55x SRAM is in fact still 1.5x density across the chip. The issue with N3 is you're getting that same 1.5x logic, but only 1.05x SRAM, so balanced across the chip it's only 1.3x.
Though on the note of pitches, much of the density gains with 3nm are from DTCO-centric changes to standard cell layout, rather than reducing pitches. Which is also why SRAM isn't scaling, because it doesn't have extra fins to cut out.
I don't think it did... most of the time TSMC was using logic density improvements in their marketing numbers.
That's how N3 vs N5 went from 1.7x to ~1.3x density.
Which is my point.
People are surprised to see the node's advertised density scaling going from the expected ~1.6x-1.8x down to 1.1x when it isn't apples to apples.
https://www.anandtech.com/show/1735...n3e-in-2024-n2-in-2026-major-changes-incoming
Last edited:
Their marketing included both logic and SRAM, separately. Whether Anandtech included SRAM cell size that in their charts is a different matter.
But what I really meant was that up until and including N7, SRAM scaling was in line or better than logic. Advertising "logic density" or "whole chip density" as multiples would give very similar numbers, it's only N5 and below where the difference has been significant.
But I did forget that TSMC had formerly given logic and SRAM separately for N5 and N3 before switching to the whole chip metric
7nm to 5nm = 1.8x density
5nm to 3nm is only 1.3x density
3nm to 2nm is only 1.15x
2nm to 1.6nm (A16) is only 1.10x.
That is a very very worrying trend. Cache density almost reach point of non-scaling, which can be parttially ofset by stacking or chiplets. Anyway keeping Moore law alive while making new gen chips are going to be tough
TSMC states that 'N2 development is well on track high-volume manufacturing in the second half of 2025.and N2P is next
also TSMC outlined a roadmap that will take the technology from a current bump pitch of 9μm all the way down to a 3μm pitch by 2027, stacking together combinations of A16 and N2 dies
https://www.anandtech.com/show/2141...aging-making-quick-progress-3um-pitch-in-2027
5nm to 3nm is only 1.3x density
3nm to 2nm is only 1.15x
2nm to 1.6nm (A16) is only 1.10x.
That is a very very worrying trend. Cache density almost reach point of non-scaling, which can be parttially ofset by stacking or chiplets. Anyway keeping Moore law alive while making new gen chips are going to be tough
TSMC states that 'N2 development is well on track high-volume manufacturing in the second half of 2025.and N2P is next
also TSMC outlined a roadmap that will take the technology from a current bump pitch of 9μm all the way down to a 3μm pitch by 2027, stacking together combinations of A16 and N2 dies
https://www.anandtech.com/show/2141...aging-making-quick-progress-3um-pitch-in-2027
Last edited:
https://wccftech.com/tsmc-yields-will-receive-high-na-chipmaking-machine-this-year-report/
interesting twist
TSMC Yields & Will Receive High NA Chipmaking Machine This Year – Report
interesting twist
There were rumors(speculation?) that TSMC had already gotten one.https://wccftech.com/tsmc-yields-will-receive-high-na-chipmaking-machine-this-year-report/
TSMC Yields & Will Receive High NA Chipmaking Machine This Year – Report
interesting twist
Given how precious they are, I dont understand how any new agreement could just come up for a delivery this year just after a recent meeting. Surely orders for every single one of these machines in the next couple years would have been set in stone well prior. Intel seems pretty 'all in' on it, so will definitely want every machine they can get to get scaled up ASAP. Maybe they're waiting for the higher throughput EXE5200 machines?
Similar threads
