Semiwiki on LinkedIn Semiwiki on Facebook Semiwiki on Twitter Register / Log in 
Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/nvidia-in-talks-to-use-intel-as-a-foundry-to-manufacture-chips.15694/
        )

    [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
)

Nvidia in Talks to Use Intel as a Foundry to Manufacture Chips

Very misleading title...

"Huang also said that Nvidia (NVDA) would be interested in using Intel (NASDAQ:INTC) as a foundry, building on its existing relationship with the company. He added that any potential use of Intel (INTC) as a foundry is not likely to happen soon, however."

Nvidia tries to use Samsung but keeps getting low yields and delays so they will no doubt try Intel if 18A pans out but it needs to be competitive with TSMC in PPAC. Nvidia is at TSMC for N4 and N3E.

Seriously, we all want Intel Foundry to be successful, it is in Pat's capable hands, absolutely.
 
  • Like
Reactions: VCT
So Nvidia will use Intel's fabs while Intel will be making their own GPUs at TSMC?

I am simple man, can anyone explain how is this economically viable? Where is IDM advantage? What if there will be spike in demand (will Intel limit slots for own CPUs to supply foundry customers)?
 
Jozo035 said:
So Nvidia will use Intel's fabs while Intel will be making their own GPUs at TSMC?

I am simple man, can anyone explain how is this economically viable? Where is IDM advantage? What if there will be spike in demand (will Intel limit slots for own CPUs to supply foundry customers)?
Click to expand...

Nvidia uses Samsung to keep TSMC humble and get better terms. They will do the same now with Intel and Samsung.

One thing you have to understand is the the Intel "4 nodes in 5 years" plan is very likely because it really is only 2 nodes, Intel 4nm and 20A. Intel 3nm is an optimized version of 4nm and 18A is an optimized version of 20A. This follows the mainstream foundry business with TSMC calling N7, N6, N5, N4 and N3 nodes versus nodes plus optimized versions.

For Intel Foundry, in my opinion, Intel will first do their internal products on 20A then use 18A for high volume leading edge customers. Intel and Samsung must however be competitive in PPAC and delivery to beat TSMC on the leading edge. Intel 4nm and 3nm are behind TSMC so no big customers there.
 
Jozo035 said:
So Nvidia will use Intel's fabs while Intel will be making their own GPUs at TSMC?

I am simple man, can anyone explain how is this economically viable? Where is IDM advantage? What if there will be spike in demand (will Intel limit slots for own CPUs to supply foundry customers)?
Click to expand...
Isn't it funny and crazy?

Nvidia is trying to leverage TSMC, Samsung, and Intel against each other in order to get the best deal. But there are two limitations:

1. Walmart uses this strategy all the time to get cheapest products. But semiconductor products are much more complicated and take lengthy cooperation between foundries and fabless companies. TSMC positions themselves as everyone's foundry but it doesn't mean everyone can get into the production schedule in the same time. Relationship, trust, and long term commitment (both ways) are critical.

2. If a fabless company (such as Nvidia) tries to maintain significant order quantity at Samsung and Intel at the same time, it may make Nvidia an insignificant customer at TSMC. On the other hand, TSMC can give preference to AMD, Intel, and many other companies who have products compete against Nvidia. It's a tricky situation.
 
hist78 said:
Isn't it funny and crazy?

Nvidia is trying to leverage TSMC, Samsung, and Intel against each other in order to get the best deal. But there are two limitations:

1. Walmart uses this strategy all the time to get cheapest products. But semiconductor products are much more complicated and take lengthy cooperation between foundries and fabless companies. TSMC positions themselves as everyone's foundry but it doesn't mean everyone can get into the production schedule in the same time. Relationship, trust, and long term commitment (both ways) are critical.

2. If a fabless company (such as Nvidia) tries to maintain significant order quantity at Samsung and Intel at the same time, it may make Nvidia an insignificant customer at TSMC. On the other hand, TSMC can give preference to AMD, Intel, and many other companies who have products compete against Nvidia. It's a tricky situation.
Click to expand...

It's a business decision. If you are exclusive to TSMC you get better terms. Apple and AMD embrace this, QCOM and NVDA do not. QCOM and NVDA are OG (original gansta) fabless companies who multi source.
 
Daniel Nenni said:
It's a business decision. If you are exclusive to TSMC you get better terms. Apple and AMD embrace this, QCOM and NVDA do not. QCOM and NVDA are OG (original gansta) fabless companies who multi source.
Click to expand...
Indeed it's a business decision. There are few companies can afford to do so in large scale.
 
hist78 said:
Indeed it's a business decision. There are few companies can afford to do so in large scale.
Click to expand...

It's not easy straddling leading edge fabs. The NDAs are very invasive. Example: The same group cannot design to Samsung and TSMC. There is usually an 18 month cooling off period. It is close to cleanroom conditions. Companies like NVIDIA and QCOM can put different products into different fabs without too much problems but Apple cannot do it since they do incremental designs. AMD could probably did it splitting CPUs and GPUS but it would be very difficult to split the CPU products without contamination. TSMC is very seriously about their PDK and process IP.

And of course it comes down to trust. What fab do you trust with your IP, your product schedule (yield), your chip specs (PPA), and of course who's PDK is the most trustworthy. Intel and Samsung have a lot of ground to cover here, absolutely.
 
A challenge for Nvidia and to a lesser degree for Qualcomm is their share in TSMC's revenue. Currently they are significant but not huge. If they continuously diversify to more foundries to the degree that their share in TSMC revenue become too small, it will create an unhealthy (or weaker) relationship between them and TSMC.

Source: https://www.tomshardware.com/news/amd-becomes-tsmc-third-largest-customer
Screenshot_20220324-124913.png
 
Last edited:
Daniel Nenni said:
Very misleading title...
Click to expand...

...that's an understatement

Everyone knows that Intel would be fabricating 4nm, 5nm, even 6nm SoC’s right now, if they could achieve better than 30% yields, as Samsung has been doing it for quite some time at 35% yields.

Given that painful truth, combined with Intel’s ZERO EUV production competence, Jensen Huang will have long since retired before Pat Gelsinger’s foundry pipe dreams come true.
 
The first thing I thought when I read scotten jones recent writing is Intel Israel figured out how to automate more of the production of smaller nodes. If true they will have a monopoly but there's still the problem of the fabless ecosystem and their options.
 
I'm a big fan of Jensen Huang but I do not agree with his treatment of TSMC. It shows a true lack of character in my opinion. Jensen and Morris were good friends back in the day, Nvidia was one of TSMC's first big customers and they grew up together. See this video interview. I was there and felt the true bond these two men shared:


Yet Jensen threw TSMC under the bus on more than one occasion (40nm yield) and played Samsung against TSMC which they still do today. Had Nvidia been loyal to TSMC they would be in a better place, absolutely.
 
Daniel Nenni said:
I'm a big fan of Jensen Huang but I do not agree with his treatment of TSMC. It shows a true lack of character in my opinion. Jensen and Morris were good friends back in the day, Nvidia was one of TSMC's first big customers and they grew up together. See this video interview. I was there and felt the true bond these two men shared:


Yet Jensen threw TSMC under the bus on more than one occasion (40nm yield) and played Samsung against TSMC which they still do today. Had Nvidia been loyal to TSMC they would be in a better place, absolutely.
Click to expand...
I believe both TSMC and Nvidia treasure their relationship and coorporations. But Nvidia only contributed 2.83% (see the table I posted previously) of TSMC revenue in 2021. It's hard for TSMC to treat Nvidia the same way as Apple, Mediatek, AMD, Qualcomm, and Broadcom. Nvidia might need to adjust their expection to reflect this reality.

On the other hand Nvidia is probably the largest or second largest (after Qualcomm) Samsung foundry's customer, if we exclude Samsung itself. The price and priority Samsung offered to Nvidia are too good to pass.

Fabless companies use multi sourcing to get better pricing and reduce risks. TSMC uses multi clients to pick the best deals and to spread out risks. It's an interesting real life drama to watch.
 
Last edited:
hist78 said:
I believe both TSMC and Nvidia treasure their relationship and coorporations. But Nvidia only contributed 2.83% (see the table I posted previously) of TSMC revenue in 2021. It's hard for TSMC to treat Nvidia the same way as Apple, Mediatek, AMD, Qualcomm, and Broadcom received. Nvidia might need to adjust their expection to reflect this reality.

On the other hand Nvidia is probably the largest or second largest (after Qualcomm) Samsung foundry's customer, if we exclude Samsung itself. The price and priority Samsung offered to Nvidia are too good to pass.

Fabless companies use multi sourcing to get better pricing and reduce risks. TSMC use multi clients to pick the best deals and to spread out risks. It's an interesting real life drama to watch.
Click to expand...

It is a real life drama, and if Samsung came even close to TSMC in regards to PPA and delivery I think it would be a much more even market split. Before FinFETs chip companies routinely did 2nd, 3rd, and even fourth sourcing of the same design. Today they have to chose their partners more carefully.

Samsung really botched 5nm, 4nm, and 3nm. TSMC is rapidly increasing their N5/4/3 capacity to pick up the slack. The question is will Samsung recover at GAA 2nm with both TSMC and Intel fighting it out? Serious real life drama.

Remember, Intel evaluated both TSMC and Samsung for 3nm and they chose TSMC, which was a very good call. N3 for TSMC will be HUGE!
 
Nvidia probably decided 8nm order in 2018-2019. At that time Nvidia and Samsung did not expect low yield of 8nm.
 
Just curious - have any (major) fabless customers ever been sued or had to pay fines for breach of contract with respect to IP issues between multiple foundries?
 
Daniel Nenni said:
It's not easy straddling leading edge fabs. The NDAs are very invasive. Example: The same group cannot design to Samsung and TSMC. There is usually an 18 month cooling off period. It is close to cleanroom conditions. Companies like NVIDIA and QCOM can put different products into different fabs without too much problems but Apple cannot do it since they do incremental designs. AMD could probably did it splitting CPUs and GPUS but it would be very difficult to split the CPU products without contamination. TSMC is very seriously about their PDK and process IP.

And of course it comes down to trust. What fab do you trust with your IP, your product schedule (yield), your chip specs (PPA), and of course who's PDK is the most trustworthy. Intel and Samsung have a lot of ground to cover here, absolutely.
Click to expand...
Just a couple of days ago, nVidia announced their CPU efforts for the datacenter. Intel is expected to announce the availability of their GPUs to the public any day now. So at this point there is not ONE product that nvidia has and Intel does not have a DIRECT competitive product. NOT ONE. So using your own words, how probable it is that nVidia is going to give their IP, product schedule, chip specs to Intel? Samsung is an IDM as well but as its foundry plans have expanded its LSI unit has gradually shrinked - and also Samsung foundry builds a customer base by being a buyer of the products from the likes of QCOM, nVidia and even AMD (chips and IP) for products that the mother company is producing. Is Intel going to follow a similar path? I believe that at this moment it is good talk when there is a semiconductor shortage to tell your shareholders and clients "sure we are going to consider other options" and also when you are a company that is trying to take some subsidies and contracts to state that you will consider supporting american manufacturing. But beyond that, I do not see much substance.
 
hist78 said:
A challenge for Nvidia and to a lesser degree for Qualcomm is their share in TSMC's revenue. Currently they are significant but not huge. If they continuously diversify to more foundries to the degree that their share in TSMC revenue become too small, it will create an unhealthy (or weaker) relationship between them and TSMC.

Source: https://www.tomshardware.com/news/amd-becomes-tsmc-third-largest-customer
Click to expand...

Wow, this shows us many things. Nvidia contributes a small portion of TSMC revenue, but they're now the world's 2nd Fabless company. That shows how Fabless with strong software ecosystem.
Meanwhile, TSMC capacity is required for a long time to keep NVIDIA software ecosystem competitive since NVIDIA's leading-edge AI devices (A100, H100...etc) are relying on TSMC.
 
diediealldie said:
Wow, this shows us many things. Nvidia contributes a small portion of TSMC revenue, but they're now the world's 2nd Fabless company. That shows how Fabless with strong software ecosystem.
Meanwhile, TSMC capacity is required for a long time to keep NVIDIA software ecosystem competitive since NVIDIA's leading-edge AI devices (A100, H100...etc) are relying on TSMC.
Click to expand...
Relationship, trust, and cooperations are critical in this fabless-foundry ecosystem. A big mouth and arrogance won't help.
 

Nvidia in Talks to Use Intel as a Foundry to Manufacture Chips​

I spent quite a bit of time digging in on this question last week and I'm sorry but Jensen is doing his signature move, head fake for future negotiations with TSMC. Unfortunately for Jensen TSMC is the only leading edge game in town for his big fat chips.
 
You must log in or register to reply here.
Back
Top