20k wafers (perhaps 450mm ones?) with unknown yield (likely 80ish) is quite low. For reference AMD 5700 has die area of ~250mm.
20k wafers (perhaps 450mm ones?) with unknown yield (likely 80ish) is quite low. For reference AMD 5700 has die area of ~250mm.
If the fab is located inside the US, AMD and NVIDIA can bid for many US government contracts that would go for Intel. After GlobalFoundries quit the race, US defense contractors have only Intel as supplier for many components.
If the consumer product is a toy then sure. If it’s a medical device then it could make sense. It’s all about margin.
> After GlobalFoundries quit the race, US defense contractors have only Intel as supplier for many components.
Maybe. If you want millions of chips a quarter then sure. But if you want 1000s a quarter then there are many options, depending on the node I guess. I’m not going to Google a complete list but one obvious example is Texas Instruments.
Small run 22nm is certainly available. These are fabricated in the US, albeit at an Intel site. https://www.mosis.com/pages/intel_sponsored_fab
What process node Google uses for TPU, one possibility for an AI chip, is not in the public domain since v1 but it’s very unlikely this chip is on the bleeding edge. Cost/benefit is likely not there for this use case. But only Google knows for sure since they run the cloud and get to define “benefit”.
https://www.nextplatform.com/2018/05/10/tearing-apart-google...