Doesn't that not work if Samsung's nodes kinda suck? Samsung's 8 nm had notorious yield problems with the 30 series, and their 5 nm class nodes still lag behind TSMC's.
Which is why Nvidia has to prop them up, instead of letting the two compete on quality. If TSMC and Samsung were both in healthy places, Nvidia wouldn't be eyeballing the market nervously.
While competition among fabs is good for fabless chipmakers, I imagine Nvidia doesn't want what is likely to be one of its most popular products to be saddled with a node with lower profitability and especially lower power efficiency.
But it isn't less profitable. That's my point. If Samsung makes the contract based on functional dies instead of wafers, from Nvidia's point of view, the yields are irrelevant, because Samsung is eating the cost. From Samsung's point of view, the node is running under capacity. They're already eating the cost, keeping the fab online. The cost of the physical wafer is a drop in the bucket.
In the case of RTX 30, Nvidia was willing to put out a less power efficient product, if Samsung would make the cost worth it, because TSMC was being extortionist. The 30 series outperformed AMD's offering anyway, Samsung lived another generation, and Nvidia made a lot of profit. The Extremely Online Gamer Class whinged about Nvidia not giving them
In consumer graphics cards and Orin, power efficiency is less paramount because they either have an external power supply where the cost is more or less invisible, or they're on devices with huge batteries. In the high end data center chips, power efficiency is much more important because their clientele can see exactly how much running a given chip costs them and compare to other chips, so Nvidia made the switch to TSMC N7. On the low end of a tablet, power efficiency starts to matter again since they're relying on a relatively small battery and need to satisy Nintendo's requirements. No amount of ass-kissing on Samsung's part is gonna make a square peg fit in a round hole.
Which is exactly what I meant when I said "I don't think 8nm is viable." At some point Nvidia and Nintendo sat down and hashed out requirements - on performance, on features, on power draw, size, cost, everything. Nvidia almost definitely offered a menu of options to Nintendo. That menu wasn't a set of nodes, but high level chip designs, with estimates for all of these parameters. An 8nm based chip was almost definitely one of the options - it would be insane if it wasn't.
I tend to believe that a chip Drake's size would never meet Nintendo's likely battery life goals on 8nm. That's me assuming a lot, not just about the their battery life goals, but the ultimate size and shape of the final hardware. That's also me assuming that there aren't major power savings that Nvidia could provide when customizing for Nintendo's use case.
If Nvidia figures out how to offer 5nm class power/performance on an 8nm node, by highly customizing to Nintendo's use case...
fuck yes. That would only be worth doing if Samsung offers
significant cost savings. I'm
here for a cheaper Switch 2 that doesn't take the sorts of crazy financial risks that result in Sony and MS laying off a 20% of their workforce. I'm absolutely into
even more amazing custom engineering even if some people think it's less sexy than "we just bought the better node."
I'm betting against that, honestly. Because I don't see where those savings could come from, and why they wouldn't have appeared in other Nvidia products. But if it happens, then, I'm more fascinated from a technical perspective than I was before.