Unlikely, but Nvidia's next gaming GPUs could be built using Intel's 18A process

[zohaibahd]

Rumor mill: A new analyst report suggests that Nvidia may use Intel's foundries to manufacture some of its future gaming GPUs. If true, this would mark a significant win for Intel's foundry business as it aims to gain market share from TSMC.

The information comes from Timothy Arcuri, an analyst at Swiss investment bank UBS. He stated in a press note to investors that Nvidia is "closer" than Broadcom to utilizing Intel's 18A process node for at least some of its graphics processors targeted at gamers.

Intel's 18A refers to its 1.8nm-class transistor technology, currently in development. A low-power variant, 18AP, is also expected, which would appeal to chip companies with power-efficient needs.

If Nvidia does opt for 18A or 18AP for future gaming products, it would likely serve as a major endorsement of Intel's process capabilities. The chipmaker has faced significant challenges in recent years trying to keep pace with TSMC on the leading edge. However, securing a client like Nvidia could signal that Intel has finally regained its competitive edge.

UBS analyst Timothy Arcuri stated, "Intel's (INTC) new CEO Lip-Bu Tan is likely to refocus on chip design in the short term and seek to attract major customers to strengthen the foundry business."

He further analyzed, "In the short term, efforts will emphasize Intel's design and…

– Jukanlosreve (@Jukanlosreve) March 25, 2025

Arcuri also suggests that the newly appointed Intel CEO, Lip-Bu Tan, will focus on chip design in the short term and aim to attract major clients like Nvidia, potentially shifting some of their production away from TSMC and Samsung's foundries.

Tan, the former CEO of Cadence Design Systems, was appointed as Intel's permanent CEO earlier this month. Securing major customers appears to be a key focus for him as he works to revitalize Intel's struggling division.

The note further adds that Intel could leverage packaging technologies such as its Embedded Multi-die Interconnect Bridge to compete with TSMC's successful Chip on Wafer on Substrate solutions.

It also mentions that Intel is already working on a partnership with Taiwan's contract chipmaker, United Microelectronics Corporation. Arcuri speculates that this deal could accelerate, with the two companies potentially co-manufacturing some of Apple's chips on Intel's high-voltage FinFET nodes as early as next year.

Of course, all of this remains in the realm of rumors for now. However, Arcuri promises that more concrete updates on Intel's foundry progress will be revealed at the company's Direct Connect event on April 29.

Permalink to story:

Unlikely, but Nvidia's next gaming GPUs could be built using Intel's 18A process

 

[GodisanAtheist]

Frankly I'd love to see AMD/Nvidia/et al. use a broad range of fabs to help satisfy demand for their chips. Sort of like when AMD was using GloFo for their Polaris chips while using TSMC for Vega.

No, the fabs are not going to all output at the same quality, but simply diversifying and broadening the product offerings on tap can both drive revenue and reduce cost to the end user if done in a smart way.

 

[Theinsanegamer]

Frankly I'd love to see AMD/Nvidia/et al. use a broad range of fabs to help satisfy demand for their chips. Sort of like when AMD was using GloFo for their Polaris chips while using TSMC for Vega.

No, the fabs are not going to all output at the same quality, but simply diversifying and broadening the product offerings on tap can both drive revenue and reduce cost to the end user if done in a smart way.

It costs way more to port a design to multiple fabs, polaris and vega were two separate arches. Even when NVidia did it, it was only on very low end parts, like the GT series or the 1050/ti.

 

[GodisanAtheist]

It costs way more to port a design to multiple fabs, polaris and vega were two separate arches. Even when NVidia did it, it was only on very low end parts, like the GT series or the 1050/ti.

- If AMD could do it while they were dummy broke back in the Bulldozer era, they could definitely do it now. Nvidia can do whatever the hell it wants it has **** you money (like flipping TSMC the finger and fabbing the whole Ampere line on Samsung's N8 process).

Even now AMD left the RX 7600 (and the MCDs on the chiplet parts) back on TSMC's N6 process while the GCDs of the RX 7800/7900 were fabbed on the more cutting edge N5.

Samsung/Intel would undoubtedly hand out some sweetheart deals to get 3rd parties to use their fab capacity now that their own internal lines of business can't keep them occupied.

 

[p51d007]

I just hope TSMC is able to broaden it's footprint OUTSIDE of Taiwan. You have to know that one day soon, China is going to cut them off from the outside world. Of course, if that happens, I will probably trigger something none of us want.

 

[Theinsanegamer]

- If AMD could do it while they were dummy broke back in the Bulldozer era, they could definitely do it now. Nvidia can do whatever the hell it wants it has **** you money (like flipping TSMC the finger and fabbing the whole Ampere line on Samsung's N8 process).

Ampere isnt comparable. They designed Ampere to run on Samsung 8nm.

Polaris and Vega were two totally different products on two different production lines, not the same product on two different lines. Nvidia has done it, again, with the 1050/ti and the 750/ti, both low end small products produced in large quantities, they were not making 1080tis on Samsung's node, and crucially, the 1050/ti were not made on TSMC nodes either, the arch was tweaked to run on the samsung tech

And I dont think that AMD's strategy, at a time they were loosing money and facing imminent bankruptcy, is a very good place to pull strategy from.

From a historical perspective, we've also seen the bad side of doing this. Remember Rocket lake and Tiger lake? Same arch, but Rocket lake was backported to 14nm and as a result, it had to have its L2 cache cut in half which kneecapped performance. Now, I want you to imagine how, on an even more complex chip (RTXC 5090) the problems that would arise from going from TSMC 3N to Intel 18a or Samsung 3nm, which were totally different products from other companies.

This isnt like a car part made of steel where anyone with access to a tool and die maker could copy the part. Engineering an architecture to run on tow or more totally unrelated production nodes from two different companies is very complex AND very expensive, for little perceived benefit to the host, especially when TSMC is currently nuking the competition from orbit. The resulting frankenstein design would either be restricted to older TSMC nodes or would need to give up significant performance design to maintain compatibility. If it were so easy to do without consequences, dont you think nVidia would have doen it with their AI GPUs already?

Even now AMD left the RX 7600 (and the MCDs on the chiplet parts) back on TSMC's N6 process while the GCDs of the RX 7800/7900 were fabbed on the more cutting edge N5.

Samsung/Intel would undoubtedly hand out some sweetheart deals to get 3rd parties to use their fab capacity now that their own internal lines of business can't keep them occupied.

TSMC 6n is a derivative of TSMC 5n. I dont see any products that are made on both TSMC 7n and 5n, for example, because the chip would need to be re engineered to work on one or the other.

 

[godrilla]

I'm surprised they didn't go with Samsung this time around again with such a crazy shortage.

 

[Theinsanegamer]

I'm surprised they didn't go with Samsung this time around again with such a crazy shortage.

Samsung has been struggling with yields on anything below 8nm. It's 5nm tech wasn't ready for customer samples until june 2023, 8 months AFTER the RTX 4080's release, and it's yields still pale compared to TSMC. Nvidia probably figured that since Blackwell really isnt all that different it'd be cheaper to use the same node again instead of moving back to Samsung.

I mean, nVidia made record revenue from their datacenter and the geforce division is holding strong at over $10 billion a year. The shortage keeps their products in demand with no price cuts for their entire lifetime, which means more $$$ for nvidia.

 

[GodisanAtheist]

And I dont think that AMD's strategy, at a time they were loosing money and facing imminent bankruptcy, is a very good place to pull strategy from.

- My point was that AMD could afford to do it when they were on the verge of bankruptcy (granted there were some extenuating circumstances, such as the wafer supply agreement) then they could absolutely afford to do it now.

I think the market is hungry for sweet spot products. If AMD (or NV but who are we kidding) could get Samsung/Intel to fab something small and reasonably performant that could make money at $200 price point, they could be the hero of the people by just stamping those things out.at an alt foundry. Samsung's foundry is in such a bad way they'd likely rebate/write off poor bins or throw some really sweetheart stuff into that fan contract.

TSMC 6n is a derivative of TSMC 5n. I dont see any products that are made on both TSMC 7n and 5n, for example, because the chip would need to be re engineered to work on one or the other.

- 6N is a derivative of 7N, 5N is a new node and 4N (which N48 is fabbed on) is a derivative of 5N. AMD is making GPUs essentially on both 7N (6N) and 5N in the 7xxx series.

 

[Vanderlinde]

Even now AMD left the RX 7600 (and the MCDs on the chiplet parts) back on TSMC's N6 process while the GCDs of the RX 7800/7900 were fabbed on the more cutting edge N5.

You do know that MCD's or Chiplet parts don't need cutting edge tech right?

 

[GodisanAtheist]

You do know that MCD's or Chiplet parts don't need cutting edge tech right?

-Yeah what does that have to do with anything? RX7600 was fabbed on N6 while the GCD of the RX7800/7900 was fabbed on N5.

So AMD already has two different processes in one generation.

The chiplet thing takes it a step further with them having two different processes on one package.

 

[Axeia]

If NVIDIA still cares I could see them splitting server (AI) products and gaming products. Keep TSMC for AI and instead of selling the gaming market scraps as they currently are have Intel dedicated itself to it.
Not like gaming needs to be super power efficiënt. As long as it clocks high enough and yields/costs are acceptable Intels process might be a good fit. The gaming market used to be NVIDIAs only market so dedicating resources to it shouldn't be that hard.

Splitting the architecture just a tiny bit might also prevent people from buying **90 class cards for AI. Obviously those cheapskates should buy much more expensive AI cards, Jenssen can barely afford more leather jackets.

I'd really like to see AMD CPUs made in Intel fabs. It'd be hilarious - heck even if they just have the IO tile made at Intel it'd be pretty funny. We've already had Intel CPUs with AMD graphics on the same chip so why not.
We've already seen