I really don't want to get into process node posting, but either this chart's numbers are wrong, or the entire discussion where "Orin's power consumption is bad on 8nm and we need to move to at least 5nm for it to be better" has been foundationally flawed. It's not clear to me which node Orin's 8N is based on, but let's assume it's 8LPP for now. According to this chart, 8LPU offers a whopping 60% reduction in power consumption over 8LPP. That already makes it
better than 5LPP! 5LPP would have 46.8% the power consumption of 10LPP, as you calculated, while 8LPU would have only 36% of 10LPP. 8LPA would be even better at 30.6% of 10LPP.
According to the other Samsung chart in
this post, 8LPU is the direct evolution of 8LPP. So if T239 needed to move off 8N/8LPP to improve power consumption, why would they make the more expensive and difficult jump to branch of the family with 5LPP, even though 5LPP is
less efficient than the directly available 8LPU? If 8N is actually based on 8LPU instead (which
this article claims is the case), then this would make even less than zero sense, since moving from 8N/8LPU to 5LPP would actually result in a less power efficient chip than if they had just stuck with what Orin was already using.
If we believe the chart, the first possible node that would result in power savings compared to 8LPU would be 4LPP, at 25.2% of 10LPP's power consumption. Although, that's if you even trust that number, since it's based on Samsung's claim from
sometime last year about a node that didn't exist then and possibly still doesn't exist today.
So I'm wondering if this chart is just totally wrong. It seems to be based entirely on Samsung's marketing claims and I don't see an indication that any testing was done to verify it.
I feel the need to state explicitly that even if the numbers are all accurate, none of this means a worse performing node is going to be used. It would just mean the 5nm rumor is wrong, and things go back to being between Samsung 8nm (though potentially a significantly better 8nm than Orin uses) and TSMC. And with that, I will try to again start ignoring process node discussions because I
really just don't care.