Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/different-estimates-of-euv-tool-allocation-at-tsmc-and-samsung-but-same-trends.15232/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2021370
            [XFI] => 1050270
        )

    [wordpress] => /var/www/html
)

Different estimates of EUV tool allocation at TSMC and Samsung but same trends

Fred Chen

Moderator
CTEE and Mizuho gave different estimates of EUV tools allocated at TSMC and Samsung but the trends are the same:
Estimates of EUV tools among customers.png

CTEE: https://ctee.com.tw/news/stocks/564615.html
Mizuho: https://seekingalpha.com/article/4424007-asml-stock-in-sweet-spot

Notably, Intel (part of the "Others") would not be able to feasibly catch up (to TSMC) in EUV volume.
 
What going on with the obvious drop in "Others" - you would have assumed that these numbers would only increase over the years ? Was that Intel disposing of some early EUV machines ? Or someone else ?
 
What going on with the obvious drop in "Others" - you would have assumed that these numbers would only increase over the years ? Was that Intel disposing of some early EUV machines ? Or someone else ?
I think Intel only bought a few of them each year in the past. In 2020 they might even cancel a 2-unit order.
 
Last edited:
  • Like
Reactions: VCT
Memory makers are buying EUV tools now? Are they in the labs or on the production floor? Will Intel have enough for their "Intel Accelerated" roadmap?

2H-2022 – 4nm (formerly 7nm) – 20% performance per watt improvement, first EUV use at Intel.
2H-2023 – 3nm – 18% performance improvement, denser libraries (some density increase) and more EUV use.
1H-2024 – 20A (2nm) – RibbonFET (first generation Intel GAA), a type of horizontal nanosheet.
1H-2025 – 18A (1.8nm) – RibbonFET (second generation Intel GAA).
 
Supposedly Intel had a (at least) 15-unit order under Brian Krzanich announced in April 2015, not sure how it was distributed over the years.

It's very possible the customer mentioned in the ASML announcement was Intel. But is there any possibility the unnamed company was Globalfundries?
 
What going on with the obvious drop in "Others" - you would have assumed that these numbers would only increase over the years ? Was that Intel disposing of some early EUV machines ? Or someone else ?
These are tool production per year, not aggregate # of EUV tools. So "Others" going down from '18 -> '19 or '19 -> 20 depending on which chart, just signifies that of what ASML was able to produce that year, the majority went to T / S. For Samsung, also not entirely clear if that volume of tools is just for logic or also logic + DRAM. TSMC and Samsung both started mass production of technologies (N7+ and S7LPP) in 2019, but the demands for N5 in terms of # of EUV layers is higher than for S5, and TSMC now produces both N7+ and N6 nodes, concurrent to N5 at different factories, all of which consume many EUV layers and thus tools. Samsung OTOH only produced maybe 2-3 designs on S7LPP before introducing S5LPE, and these are effectively the same node / same factory, just different std cell library used in the latter.
 
CTEE is not a credible source.
Most reporters left the company. Current reporters mostly copy news on the Internet without further investigation.
 
It's very possible the customer mentioned in the ASML announcement was Intel. But is there any possibility the unnamed company was Globalfundries?
Don't think so. GloFo decided not to join the EUV arms race already. Since DRAM makers were far away from mass EUV adoption, that must be Intel.
 
Memory makers are buying EUV tools now? Are they in the labs or on the production floor? Will Intel have enough for their "Intel Accelerated" roadmap?

2H-2022 – 4nm (formerly 7nm) – 20% performance per watt improvement, first EUV use at Intel.
2H-2023 – 3nm – 18% performance improvement, denser libraries (some density increase) and more EUV use.
1H-2024 – 20A (2nm) – RibbonFET (first generation Intel GAA), a type of horizontal nanosheet.
1H-2025 – 18A (1.8nm) – RibbonFET (second generation Intel GAA).
Both Samsung and SK Hynix say they are/will use EUV in their newest DRAM nodes. (Micron chose "computational lithography" instead, no EUV)

Samsung also says they are in high volume manufacturing on that EUV DRAM node - https://news.samsung.com/global/samsung-starts-mass-production-of-most-advanced-14nm-euv-ddr5-dram .
Due to this I feel most EUV machine allocation estimates are not correct when it comes to SK Hynix especially.
 
Both Samsung and SK Hynix say they are/will use EUV in their newest DRAM nodes. (Micron chose "computational lithography" instead, no EUV)

Samsung also says they are in high volume manufacturing on that EUV DRAM node - https://news.samsung.com/global/samsung-starts-mass-production-of-most-advanced-14nm-euv-ddr5-dram .
Due to this I feel most EUV machine allocation estimates are not correct when it comes to SK Hynix especially.
Samsung's allocation between DRAM and foundry is still unknown; perhaps that is undercounted. TSMC's portion is still expected to be the majority, based on the number of fabs and layers where EUV is allocated. With the total known, and a reasonable range of estimates for TSMC, Samsung, and even Intel, very few are left. SK Hynix was mentioned to have two by end of this year: https://english.etnews.com/20210629200001
 
Memory makers are buying EUV tools now? Are they in the labs or on the production floor?

This was announced Oct 2021, “begun mass producing”. Samsung Foundry fabs also produce DRAM.
 

This was announced Oct 2021, “begun mass producing”. Samsung Foundry fabs also produce DRAM.
For Samsung, the same fab (being very large) can be used to do NAND, DRAM, Foundry independently.
 
Notably, Intel (part of the "Others") would not be able to feasibly catch up (to TSMC) in EUV volume.

If Pat negotiated an addendum to Bob Swan’s original 3nm order, hopefully TSMC wrestled a few EUV machines away from Intel’s future allotment.
 

Samsung 5nm (5LPE) analyzed by TechInsights to have 8 EUV layers: CT, M0-3,V0-2

TSMC would have extra mask exposures for cuts as well as some double patterning (just for 5nm). So they could easily demand at least 2x however many Samsung uses.

Samsung could be using double patterning as well for 4nm, so their mask exposures would also go up.

For 150000 wpm, 3 fab phases, this amounts to <70 wph (yield, throughput, utilization). So it does look like low yield at advertised throughput.
 
Back
Top