NVIDIA Once Again Teases Potential Deal With Intel: Says Would Love To Have A Third Foundry Partner Besides TSMC & Samsung

[Daniel Nenni]

NVIDIA's Chief Financial Officer Colette Kress has once again teased a potential partnership with Intel Foundry Services for next-gen chip production.

NVIDIA Calls TSMC A Great Partner, Samsung Still In Use To This Day & Intel A Potential "Third" Foundry Partner

During the most recent UBS Global Technology Conference, NVIDIA's CFO was asked if they would consider going to Intel as a Foundry partner for the production of their next-gen chips. To this, the CFO showed very positive remarks based on their statement.

The bulk of NVIDIA's Data Center GPUs for AI/HPC and Gaming chips are currently being produced by TSMC but just one generation ago, Samsung was the one building NVIDIA's Gaming GPUs. The Samsung Fabs were responsible for the development of Ampere, NVIDIA's GPU family that powers the GeForce RTX 30 "Gaming" graphics cards. However, as extensively reported in our previous coverage, Samsung is aiming to oversee a much more crucial role in the "nourishing" of NVIDIA's data center revenue for the coming years, since the firm has reached a position that has made it a close partner to Team Green.



It is expected that TSMC will retain its key partnership with NVIDIA for the production of Hopper H200 and Blackwell B100 GPUs, retaining its AI market share momentum while Samsung will be available in case additional orders are required. NVIDIA also aims to have a rich ecosystem of foundry partners and is open to using a third one (referring to Intel). Here is what NVIDIA had to say:

I think there are a lot of great foundry partners. TSMC has been a great one. As you know, we also use Samsung today. Would we love a third one? Sure. We would love a third one. And that takes a work of what are they interested in terms of the services. Keep in mind, there is other ones that may come to the U.S. TSMC in the U.S. may be an option for us as well. Not necessarily different, but again in terms of the different region. Nothing that stops us from potentially adding another foundry.
Colette Kress (NVIDIA CFO) at UBS Global Technology Conference

Samsung Foundry has framed its offerings in a way that has compelled team green to increase its order placement volumes to the firm, given that not only has the Korean giant refined its processes, but has continuously collaborated with Team Green for the past two years to become a reliable supplier.

The edge Samsung has over other foundry competitors is its extensive line of equipment that can cater to semiconductor, memory, and packaging stages, which not only provides NVIDIA a "hybrid partner" but Samsung could potentially give itself a boost if the company offers competitive pricing and delivery times.



Industry sources have mentioned the name of Intel as a foundry partner. Even NVIDIA's CEO, Jensen Huang hinted at using Intel's Foundries for upcoming chips. A deal with NVIDIA would definitely be a major deal for Intel as the company aims to gain more IFS partners in the future.

Having US-based Foundry partners will also be beneficial for NVIDIA and other chip giants such as AMD and even Intel which relies on TSMC for the production of certain IPs in the chips. But NVIDIA CEO believes that the US is still decades away from supply chain independence and bringing TSMC and other major fabs to the US will be the first major step in achieving this goal.

NVIDIA Once Again Teases Potential Deal With Intel: Says Would Love To Have A Third Foundry Partner Besides TSMC & Samsung

NVIDIA Chief Financial Officer Colette Kress has once again teased a potential partnership with Intel Foundry Services for next-gen chips.

wccftech.com

 

[Daniel Nenni]

With the advent of chiplets, big chip companies like NVDA, AMD, QCOM, etc... are much more likely to use multiple foundries. Intel already uses TSMC for chiplets so 2nd and 3rd source manufacturing is coming back to the semiconductor industry, absolutely. The good news is that chiplets will spur innovation/competition in both semiconductor process technology and packaging, for the greater good of the semiconductor industry, absolutely.

 

[blueone]

If I really thought Nvidia was going to contract with IFS anytime soon, I'd be buying 100K 9/20/2024 INTC $60 calls for $1.08 each (last price). Nah. I know, no guts no glory, but it still feels like too much of a stretch until IFS proves itself with an external customer.

 

[Daniel Nenni]

If I really thought Nvidia was going to contract with IFS anytime soon, I'd be buying 100K 9/20/2024 INTC $60 calls for $1.08 each (last price). Nah. I know, no guts no glory, but it still feels like too much of a stretch until IFS proves itself with an external customer.

Understood. I'm sure IFS customer press releases are in process but it will be YEARS before IFS sees significant revenue from 18A customers. That is the proof we all need is customer success, not just a whale of a press release.

 

[cliff]

If I really thought Nvidia was going to contract with IFS anytime soon, I'd be buying 100K 9/20/2024 INTC $60 calls for $1.08 each (last price). Nah. I know, no guts no glory, but it still feels like too much of a stretch until IFS proves itself with an external customer.

Mr. Blue, doesn't Intel already create SOCs on these processes? Do you think the yield is terrible? What is stopping Nvidia from using Intel's foundry?

Why is it unrealistic for a large company to want to design advanced chips with packaging in the US? China cut internet cables to Taiwan in March. They dragged an anchor for 20 miles near Finland in October to cut cables. The Taiwan elections are going to take place in January. It seems to me that 2nd/3rd sourcing at Intel is a no-brainer.

 

[blueone]

Mr. Blue, doesn't Intel already create SOCs on these processes?

Not production chips on 18A, to my knowledge.

Do you think the yield is terrible?

I don't have any information on yield.

What is stopping Nvidia from using Intel's foundry?

Intel's readiness to be an external customer-oriented fab, with proven organization processes, tool readiness, and cooperation skills. I think this stuff, among other critical readiness factors, take time to develop, optimize, and mature.

Why is it unrealistic for a large company to want to design advanced chips with packaging in the US? China cut internet cables to Taiwan in March. They dragged an anchor for 20 miles near Finland in October to cut cables. The Taiwan elections are going to take place in January. It seems to me that 2nd/3rd sourcing at Intel is a no-brainer.

It is not unrealistic, in fact I think there is a strong business case for developing these capabilities on US soil. It is a matter of readiness to have a frenemy business relationship with a company with four or five times your market capitalization, where excellent results will be simply table stakes to keep the customer on board.

 

[cliff]

So Intel is only shipping chips based on DUV machines?

 

[blueone]

So Intel is only shipping chips based on DUV machines?

Intel 4 uses EUV in the Ireland Fab for the Meteor Lake CPUs.

Intel’s New Fab in Ireland Begins High-Volume Production of Intel 4...

Intel deploys extreme ultraviolet lithography in high-volume manufacturing; use of the technology in Ireland is a first for mass production in Europe.

www.intel.com

But Dan and I were discussing the 18A process, which isn't in production yet.

 

[cliff]

EUV not ready. Gotcha...

and based on what you wrote, it seems that you don't believe Intel can work with external customers on their double patterned DUV processes either, which has nothing to do with the foundry, but due to immature DRC (tool readiness) and management incompetence (the rest of what you wrote). Is that correct?

 

[blueone]

EUV not ready. Gotcha...

The reference I posted said EUV is ready and in production.

and based on what you wrote, it seems that you don't believe Intel can work with external customers on their double patterned DUV processes either, which has nothing to do with the foundry, but due to immature DRC (tool readiness) and management incompetence (the rest of what you wrote). Is that correct?

Not completely correct. I didn't mention management incompetence. I think it's going to be more of a case of lack of management knowledge and/or experience with best practices in being a foundry. I wonder how many former foundry users are on staff in IFS?

 

[cliff]

OK, so IFS is independent of the foundry that Intel's CPU, etc divisions use? I assumed it was the same foundry, PDK, etc, just that Intel was going to let external companies use it.

 

[hist78]

Does Nvidia have enough quantity to spread out to three foundries and that is meaningful at each foundry?

 

[cliff]

They have a lot of money. Make designs based on the different foundries. They don't need to even tape it out. Simulate and lay it out. Pick 2 out of 3 later. The geopolitics are changing rapidly.

 

[blueone]

Does Nvidia have enough quantity to spread out to three foundries and that is meaningful at each foundry?

Yes, but the only reason I can think of do to it is to diversify away from TSMC. I think the client GPU business is in the range of 30-40 million units per year, and server GPUs are in the single digit millions (but growing fast). By comparison, the entire server CPU business is in the range of 40 million units per year. Perhaps Dan has access to some more accurate numbers. Many high-end ASICs have sales volumes in the low millions of units or less per year (100Gbps and above network switch ASICs come to mind), so while millions of anything isn't that impressive, it is in the realm of what other companies do with foundries. I wonder how much money and person-years of effort Nvidia is willing to spend on foundry diversification?

 

[Tanj]

OK, so IFS is independent of the foundry that Intel's CPU, etc divisions use? I assumed it was the same foundry, PDK, etc, just that Intel was going to let external companies use it.

When Intel tried to be a foundry last time they were not using industry standard EDA, which I heard caused problems when dealing with outside customers. It seems like those days are gone and Intel can certainly take advanced designs to TSMC using tools others use, so they probably have solved that problem. Which is not to say Blue's other concerns are not valid, just that some of them are past.

 

[Tanj]

I think it's going to be more of a case of lack of management knowledge and/or experience with best practices in being a foundry.

Look at it from Nvidia's POV. Since they are now THE WHALE they need to deal with Intel and be ready for IFS. Yes, it may be a mess now, all the more reason to get started and have a real commercial chip in process, no doubt one they can tolerate some delays, because they need to pipeclean the relationship and only a real project will do that. Waiting longer is not going to make that easier or cheaper.

 

[cliff]

I wonder how much money and person-years of effort Nvidia is willing to spend on foundry diversification?

It costs way less than you think. The testbenches and floorplans are already done. It's just about the same circuit.

We can migrate a SerDes and SARADC in a month. It takes 30 minutes to migrate the TBs and DUTs automatically, but then the optimizing/running, compaction, prerouting, autoroute, RC extraction, rinse and repeat takes a few weeks. Digital should be mostly automated. The SRAM should be compiled.

This is a no-brainer. China's been cutting cables.

 

[hist78]

They have a lot of money. Make designs based on the different foundries. They don't need to even tape it out. Simulate and lay it out. Pick 2 out of 3 later. The geopolitics are changing rapidly.

Foundries are not stupid. If Nvidia keeps doing design only projects without placing large orders or large enough orders, foundries won't reserve the capacity for Nvidia and won't treat Nvidia as important as other Nvidia's competitors.

The geopolitics do not wipe out Nvidia's competitors or companies who need foundry capacity, such as AMD, MediaTek, Qualcomm, and even Intel. They will be happy to take over the foundry capacity Nvidia doesn't commit or doesn't want to pay at certain terms.

It's especially true when the market is hot and the capacity is tight.

 

[cliff]

What does creating the design and simulating have to do with the foundry resources? Just get the damn PDK and start putting together the design with the EDA TOOLS, then verify with the EDA TOOLS.

And run the damn MPW shuttles!

 

[blueone]

It costs way less than you think. The testbenches and floorplans are already done. It's just about the same circuit.

We can migrate a SerDes and SARADC in a month. It takes 30 minutes to migrate the TBs and DUTs automatically, but then the optimizing/running, compaction, prerouting, autoroute, RC extraction, rinse and repeat takes a few weeks. Digital should be mostly automated. The SRAM should be compiled.

I believe that you can move between foundries cheaply, but Nvidia with their huge high-performance dies? I'm a skeptic.