The Intel Execution in [2025]

[homerdog]

Yes if Intel's doing their own products, but if fabbing for NVIDIA I think cost will still be an issue unless their process is much better than everyone elses', otherwise NVIDIA would just shopping around and find a more cost effective fab.
I think this is kind of a blessing and curse of vertical integration. Vertical integration allows one to overlook some cost inefficiencies in the supply chain, but on the other hand, it also might cause the supplier (which they own) becoming too inefficient because they have a guaranteed customer.
Intel is always a process first company. In a sense most of its problems can be pointed at the process: if they have a world leading process most of these problems would just disappear or at least not as serious as it is now.

I was thinking if TSMC was at capacity, NVIDIA could have Intel produce for them even if it's more expensive. The cost wouldn't be a huge deal for DC chips that sell for tens of thousands dollars.

The flaw in my logic is that these chips probably can't be ported between fabs like they are making toasters. I'd imagine even if Intel had a node that was comparable in performance to 4N, it would be a lot of work to port GB200 to it. Maybe this could change; I saw Intel was working with ASML but I dunno how long that's been going on.

 

[pcchen]

I was thinking if TSMC was at capacity, NVIDIA could have Intel produce for them even if it's more expensive. The cost wouldn't be a huge deal for DC chips that sell for tens of thousands dollars.

The flaw in my logic is that these chips probably can't be ported between fabs like they are making toasters. I'd imagine even if Intel had a node that was comparable in performance to 4N, it would be a lot of work to port GB200 to it. Maybe this could change; I saw Intel was working with ASML but I dunno how long that's been going on.

Yeah that's an issue. As other mentioned Intel's process used to use very different design rules from others. TSMC on the other hand is the industry standard for quite some time. In the past Intel didn't care about this because they basically only fab their own products, but now since they seriously want to be a foundry they are on the way to be compatible with others. How well it's another question, but at least Samsung did that to the extent that NVIDIA was willing to use Samung to fab their GPU at the same time using TSMC for their accelerators, so I think it's possible.