Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/amd-apple-nvidia-reportedly-cutting-back-on-tsmc-n5-orders-n6-n7.16309/
        )

    [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
)

AMD, Apple, Nvidia Reportedly Cutting back on TSMC N5 orders (+N6/N7)

Xebec

Well-known member
OG Source (paywall) : https://www.digitimes.com/news/a20220622PD212/3nm-ic-manufacturing-tsmc.html
TH Link: https://www.tomshardware.com/news/amd-apple-nvidia-reportedly-reducing-5nm-tsmc-orders

- Apple cutting back orders for iPhone 14 components by 10% from original 90 million units initially
- "Since Nvidia made pre-payments to TSMC for capacity and agreed to certain terms, TSMC is apparently unwilling to reduce production of chips for Nvidia, but it agreed to delay first deliveries, possibly up to Q1 2023"
- AMD reduced orders for products on TSMCs N6 and N7 process technologies by 20,000 wafers for Q4 2022 and Q1 2023.

It looks like the source article (and Toms) also talk a bit about the ramp-up of orders on the various N3 processes from TSMC.
 
OG Source (paywall) : https://www.digitimes.com/news/a20220622PD212/3nm-ic-manufacturing-tsmc.html
TH Link: https://www.tomshardware.com/news/amd-apple-nvidia-reportedly-reducing-5nm-tsmc-orders

- Apple cutting back orders for iPhone 14 components by 10% from original 90 million units initially
- "Since Nvidia made pre-payments to TSMC for capacity and agreed to certain terms, TSMC is apparently unwilling to reduce production of chips for Nvidia, but it agreed to delay first deliveries, possibly up to Q1 2023"
- AMD reduced orders for products on TSMCs N6 and N7 process technologies by 20,000 wafers for Q4 2022 and Q1 2023.

It looks like the source article (and Toms) also talk a bit about the ramp-up of orders on the various N3 processes from TSMC.

Rumor sites are fun and keeps me busy with calls from the financial community. Keep up the good work!
 
Is this the same article?


It seems like this is the original article. The article stated that the real impact on TSMC might be limited due to the demand is still probably outpacing TSMC's capacity. We will be able to find it out in about two weeks from TSMC's Q2 earnings conference call.
 
Semiconductor Hard or Soft Landing? CHIPS Act?

Robert and I are in agreement on this. My bet is over 10% semiconductor growth for 2022 and over 0% growth for 2023 (back to mid single digit growth). Since N3 revenue will start to hit next year TSMC should have another good year but back to the 20% growth versus 30% they forecasted for 2022.

Take a look at the Samsung global manufacturing sites: https://semiconductor.samsung.com/foundry/manufacturing/

Notice that Samsung 7nm and 6nm are missing? TSMC is dominating the FinFET era and their revenue surge will continue. Yes Samsung is first to GAA with 3nm but they have no big external customers, only Samsung internal chips and a couple of Chinese crypto companies and Crypto is crashing.

Intel 3 is FinFET based and competitive with TSMC N3 but the ecosystem is still ramping. Intel will definitely take the "NOT TSMC" business from Samsung but I do not see any big TSMC customers fully moving to Intel yet. Probably at 18A but 18A revenue is years away. Some of the key people who built the TSMC ecosystem now work for Intel Foundry so I can assure you TSMC is taking IFS seriously.

I have calls every week with industry experts and Wall Street. People really do not get what a strong position TSMC is in, the strongest in TSMC history, but I am happy to keep explaining it for my normal hourly rate, absolutely.
 
AMD normalized silicon area in q1 2021 at 7 nm across the full product stack ccx + i/o where required = 345 mm2. 40K wafers less q4 - q1; @ 75% yield for edge, area, scribe up to 6,147,391 completed components. If they're all Epyc which is unlikely on margin take and trading at normalized 739 mm2 = 2,869,891 completed components. What isn't showing in the channel currently is Rembrandt 6 nm mobile @ 206 mm2 = 10,295,388 components. Market reception to Rembrandt is in question on prior generation Cezanne just as performant. More so Cezanne 'new' overage and Renoir new and secondary surplus in the channel are both excessive and peaking in supply volume in parallel currently. AMD appears to be building a graded inter generation 'mobile' product heap pressing up the 'new' generation price ladder. Intel Tiger mobile overage is also excessive. Alder U/P available is slim. Alder H at 23 weeks supply is 58% of Tiger H available at 23 weeks. My take is the channel needs time to clear back generation in a first in first out sales system. Channel data does not indicate a PC slow down. Channel data indicates the channel is stuffed with back generation waiting liquidation to finance new procurements. 2019 through 2021 it was Xeon v3/v4 surplus clean up mess in the channel and now that's getting under control its mobile clean up mess in the channel. I'll update AMD full stack normalized soon. mb
OG Source (paywall) : https://www.digitimes.com/news/a20220622PD212/3nm-ic-manufacturing-tsmc.html
TH Link: https://www.tomshardware.com/news/amd-apple-nvidia-reportedly-reducing-5nm-tsmc-orders

- Apple cutting back orders for iPhone 14 components by 10% from original 90 million units initially
- "Since Nvidia made pre-payments to TSMC for capacity and agreed to certain terms, TSMC is apparently unwilling to reduce production of chips for Nvidia, but it agreed to delay first deliveries, possibly up to Q1 2023"
- AMD reduced orders for products on TSMCs N6 and N7 process technologies by 20,000 wafers for Q4 2022 and Q1 2023.

It looks like the source article (and Toms) also talk a bit about the ramp-up of orders on the various N3 processes from TSMC.

OG Source (paywall) : https://www.digitimes.com/news/a20220622PD212/3nm-ic-manufacturing-tsmc.html
TH Link: https://www.tomshardware.com/news/amd-apple-nvidia-reportedly-reducing-5nm-tsmc-orders

- Apple cutting back orders for iPhone 14 components by 10% from original 90 million units initially
- "Since Nvidia made pre-payments to TSMC for capacity and agreed to certain terms, TSMC is apparently unwilling to reduce production of chips for Nvidia, but it agreed to delay first deliveries, possibly up to Q1 2023"
- AMD reduced orders for products on TSMCs N6 and N7 process technologies by 20,000 wafers for Q4 2022 and Q1 2023.

It looks like the source article (and Toms) also talk a bit about the ramp-up of orders on the various N3 processes from TSMC.
 
Rumor sites are fun and keeps me busy with calls from the financial community. Keep up the good work!
Apologies - I thought Digitimes was generally credible.

I thought it was at least interesting that Nvidia's lack of leverage with TSMC post-20nm was showing in the "refusal to reduce production for Nvidia" portion of the article.
 
I view today's media like US politicians. Politicians spend most of their time getting re elected and saying whatever it takes to raise money etc... Most media sites earn money from clicks so they will publish whatever it takes to get those clicks. This is a personal issue for me so no offense intended. I do feel like the media is poisoning the world with misinformation all to make a profit or with some hidden agenda. It really is out of control and the world is a much worse place as a result, just my opinion of course.

For semiconductors I sit in a unique position. I'm media, but unlike most media I have 40+ years experience in the subject matter. I also have the SemiWiki ecosystem of subscribers, writers, readers, and registered members to collaborate with. The SemiWiki analytics is also quite helpful in understanding the world of semiconductors. Unfortunately I have NDAs in place that do not allow me to speak as freely as I would like but that is an important part of business. Other SemiWiki contributors have the same issue. I also have to be careful with NDAs and what is published on SemiWiki since it all comes back to me.

But, all in all, SemiWiki is still a great place to learn and network and your involvement is greatly appreciated, absolutely.
 
I thought it was at least interesting that Nvidia's lack of leverage with TSMC post-20nm was showing in the "refusal to reduce production for Nvidia" portion of the article.

Since Nvidia uses multiple foundries they are not a "preferred" TSMC customer. It may not seem fair but TSMC can more closely collaborate with dedicated customers (Apple, AMD, MediaTek, etc... ) so there are benefits not available to companies who don't (Nvidia, QCOM, Google, etc...).

The TSMC refusal to reduce Nvidia production is a rumor that I have not been able to confirm. Can anyone else confirm that? We will probably have to wait for the Q2 investor calls.
 
Since Nvidia uses multiple foundries they are not a "preferred" TSMC customer. It may not seem fair but TSMC can more closely collaborate with dedicated customers (Apple, AMD, MediaTek, etc... ) so there are benefits not available to companies who don't (Nvidia, QCOM, Google, etc...).

The TSMC refusal to reduce Nvidia production is a rumor that I have not been able to confirm. Can anyone else confirm that? We will probably have to wait for the Q2 investor calls.
Is any of the stuff in the Digitimes article confirmable? After AMD's FID just a few weeks ago it doesn't seem like they would be massively cutting orders with TSMC... especially how they very proudly emphasized doubling the size of their DC revs every quarter... I'm seeing anecdotal evidence everywhere that AMD is gaining massive share in DC this year, this comment on seeking alpha is a recent one: https://seekingalpha.com/article/45...-amd-the-king-is-coming-back#comment-92764677
 
Is any of the stuff in the Digitimes article confirmable? After AMD's FID just a few weeks ago it doesn't seem like they would be massively cutting orders with TSMC... especially how they very proudly emphasized doubling the size of their DC revs every quarter... I'm seeing anecdotal evidence everywhere that AMD is gaining massive share in DC this year, this comment on seeking alpha is a recent one: https://seekingalpha.com/article/45...-amd-the-king-is-coming-back#comment-92764677

The recent supply chain calls have Intel back in contention but AMD is still gaining. Lead times are pretty high though so I'm not seeing the slowdown on the big chips yet. These are the datacenter and HPC people. I don't track PCs and such.
 
Is any of the stuff in the Digitimes article confirmable? After AMD's FID just a few weeks ago it doesn't seem like they would be massively cutting orders with TSMC... especially how they very proudly emphasized doubling the size of their DC revs every quarter... I'm seeing anecdotal evidence everywhere that AMD is gaining massive share in DC this year, this comment on seeking alpha is a recent one: https://seekingalpha.com/article/45...-amd-the-king-is-coming-back#comment-92764677
FWIW - the article referred to AMD reducing N6 and N7 volumes, not N5. There's indication both for and against N6/N7 capacity needs:

Reduced Demand for N6/N7:

- Zen 4 is currently expected to launch in Sept-Oct meaning some of AMD's focus shifts to needing N5 capacity. This will be followed by server chips based on N5 in Q1 2023.

- AMD appears to be working to position itself as the 'premium' brand even ignoring lower end markets for up to a year at a time to avoid selling CPUs below ~ $299 MSRP, so it's possible they'll want to reduce N7 logic capacity to avoid selling too many lower margin products.

- AMD GPUs (N7) are starting to sell below MSRP due to a crypto crash. AMD may have too much inventory building up on these products especially with next gen GPUs expected in Q4 2022.

- AMD next-gen RDNA3 is based on N5 and launches in Q4. While RDNA3 is MCM and will use some N6/N7 capacity for the I/O+cache die, every RDNA3 GPU will use less N7 die space than current gen.

- The word "recession" keeps appearing in the news..

Increased Demand for N6/N7:

- Zen 4 and Epyc Genoa will require N7 I/O dies. The I/O dies have typically been pretty large though you don't need as many of them as logic dies to sell Zen based products.

- N7 / Zen 3 Server chip demand is certainly increasing

Take-away:
AMD's need for N6/N7 wafers appears to be on a downward trend but ... it's hard to say if this was in line with AMD's original capacity order and plans.
 
Nvidia and AMD ditching foundry orders? How about redirecting starts to other products. Xeon Skylake and Cascade Lake if they can be refurbished from storage and server to accelerated compute represent a 400 M unit installed base and the largest upgrade market in the history of IT. Production cost is up. AMD and Nvidia have been swinging from consumer to commercial products for four quarters. Scaling the Red Brick Wall; design, materials, process, interconnect, packaging, it's all about climbing the cost : price / margin ladder since leaving 12 nm on the now upward sloping cost curves.

I can comprehend why AMD might be reducing Rembrandt total production on the glut of Cezanne and Renoir and Intel Tiger overage, Ice and Comet mobile sitting in the channel waiting for a home. N6 for the cost increase may also NOT be as price / margin effective for the performance increase over N7. It's a good time to buy back gen 'new' laptop on clearance sale. You get what you negotiate through a broker.

Noteworthy specific laptop contract production for every AMD mobile production start Intel gets minimally 50% more starts and up to 80% more starts from contract resource back through Kaby Lake H. Intel still controls Asia Pacific mobile production on economies of production scale. To date Intel x86 mobile gets first dibs from the build resource and AMD is always second. It's easy to GANTT on public data.

Specific Nvidia, TSMC through Ampere GPU generation was evidently Nvidia's low power choice and Samsung Nvidia's frequency choice. Noteworthy and I suspect this shuck Intel, its shuck me to discover on supply data in relation prior Nvidia generations, that 50% of all Ampere A Series are mobile workstation (aka performance gaming in disguise) and in q4 2021 in excess of 60% on channel supply data. A6000 is the only pure workstation SKU. All other A Series (sans AK55/45) Professional cards include a good chuck of Mobile among their total volumes although overall A mobile is still a relatively slim quarterly volume in relation consumer mobile volume. I place Nvidia mobile @ or < 15% of Intel 'mobile quarterly compliment'. Ampere mobile represents 21.5% of channel supply volume.

AMD mobile CPU in 2021 secured around 7% mobile share and holds 15% of the channel supply back through Comet and Renoir. AMD mobile dGPU remain nil at 1% of total Navi 6x supply. In 2021 I have AMD producing 6,689,101 Epyc with a record breaking volume quarter q1 2022 up to 2,955,832 units plus a few hundred thousand thought RDNA ll accelerators not Instinct. AMD Epyc production q1 2022 just surpassed Xeon Ice production in q1.

2022 to 2025 cost : price / margin necessitates the swing to high margin commercial products with a B2B emphasis. End use (IT) ROI is king. There's plenty of consumer product in the channel that needs to clear. I don't believe AMD, Intel or Nvidia will be giving up their stronghold positions anytime soon; AMD desktop, Intel mobile & server, Nvidia dGPU, simply reconfiguring their production starts for margin gains with an emphasis on commercial and B2B. Where Intel reconfiguring from supply to demand side production is a necessity coming out of monopoly surplus 'overhang' that held channels financially on overage tied to procurement incentives moves to an emphasis on cost optimization removing unnecessary and avoidable costs.

mb
 
  • Like
Reactions: VCT
AMD lost its
FWIW - the article referred to AMD reducing N6 and N7 volumes, not N5. There's indication both for and against N6/N7 capacity needs:

Reduced Demand for N6/N7:

- Zen 4 is currently expected to launch in Sept-Oct meaning some of AMD's focus shifts to needing N5 capacity. This will be followed by server chips based on N5 in Q1 2023.

- AMD appears to be working to position itself as the 'premium' brand even ignoring lower end markets for up to a year at a time to avoid selling CPUs below ~ $299 MSRP, so it's possible they'll want to reduce N7 logic capacity to avoid selling too many lower margin products.

- AMD GPUs (N7) are starting to sell below MSRP due to a crypto crash. AMD may have too much inventory building up on these products especially with next gen GPUs expected in Q4 2022.

- AMD next-gen RDNA3 is based on N5 and launches in Q4. While RDNA3 is MCM and will use some N6/N7 capacity for the I/O+cache die, every RDNA3 GPU will use less N7 die space than current gen.

- The word "recession" keeps appearing in the news..

Increased Demand for N6/N7:

- Zen 4 and Epyc Genoa will require N7 I/O dies. The I/O dies have typically been pretty large though you don't need as many of them as logic dies to sell Zen based products.

- N7 / Zen 3 Server chip demand is certainly increasing

Take-away:
AMD's need for N6/N7 wafers appears to be on a downward trend but ... it's hard to say if this was in line with AMD's original capacity order and plans.
AMD lost its TSMC 7 nm cost advantage to Intel SF10 in q3 2021 and entered Genoa 5 nm risk (sample) production. Intel SF 10 on my production economic estimate has a 20% cost advantage over TSMC 7 primarily on the TSMC mark up to AMD and is parity cost with TSMC 5. mb
 
  • Like
Reactions: VCT
AMD mobile CPU in 2021 secured around 7% mobile share and holds 15% of the channel supply back through Comet and Renoir. AMD mobile dGPU remain nil at 1% of total Navi 6x supply. In 2021 I have AMD producing 6,689,101 Epyc with a record breaking volume quarter q1 2022 up to 2,955,832 units

You mean ~3m Epyc chiplets in 1q22? If it would be ~3m packaged chips that would work out to $9B revs at a conservative asp of $3k which doesn't make much sense to me...
 
Sailor Bob said, "You mean ~3m Epyc chiplets in 1q22?". I think AMD is still on Milan at 7 and Genoa at 5 and will be for some time. McNamara at the financial analyst day said AMD would continue to produce Milan and that makes sense if a depreciated cost 7 is competitive price / performance wise with the 400 M units installed Xeon Cascade Lake / Skylake replacement market. The general enterprise compute market, and channel, have standardized on dirt cheap Cascade Lakes platform. Cascade Lake is the generation of Xeon that Milan was designed to compete. 400 M unit installed base is a huge replacement market and channel data shows the general compute cores sweet spots; 8/10, then 12/14/16, then 20,24,28 and generally for AMD @ < 32. The general enterprise compute market is still upgrading from Haswell v3 and Skylake and Cascade Lake is where the majority of sales are.

Genoa at 5 I've thought shipping since q3 2021 because DDR5 memory required physical validation not just in simulation. At q3 2021 is when AMD lost its 7 nm cost : price / margin advantage to iSF10/x on TSMC's take of AMD revenue per unit. This had to press AMD to ramp Genoa production, to maintain price performance advantage verse Intel SF10 which justifies AMD's price while maintaining a margin target.

If hyperscale, public cloud, business of compute can code Genoa for competitive advantage they will, same with Sapphire Rapids, Instinct, Ponte Vecchio. All these products are ahead of their whole software stack waiting validation pursuant general compute market. IT always waits for proof of whole product and this is why AMD and Intel focus on hyperscale because they need the whole platform software stack validation and business of compute is in that business enabling whole product (platform) efficiencies for AMD and Intel.

3 nm? Not yet other than nascent engineering samples, maybe, in my view. 2022 is a Milan mainstream and Genoa production year is my take. I'm monitoring channel supply to confirm that thesis. It's the one that makes sense to me. mb
 
Reduced Demand for N6/N7:
- AMD GPUs (N7) are starting to sell below MSRP due to a crypto crash. AMD may have too much inventory building up on these products especially with next gen GPUs expected in Q4 2022.
Their custom products (eg: Playstation / Xbox CPUs) are also going to be subject to shifts in demand. Last two years they were extremely high, same as dGPU sales and for similar reasons (scalpers, or indirectly lack of dGPU means higher demand for consoles). So with a glut of (used, probably unstable) dGPU cards on the market, there will be the reduction of demand for new dGPU cards but may also see reduction in console demand. Alternately, they may re-spin those APUs on N6 from N7, and that scaling benefit might given them a slight reduction in total wafers required, though given the die size I doubt that. More likely wait for an N4 based refresh of the console parts and coincide with a "Pro" model launch.
 
Back
Top