Several reasons. The
TL;DR version is: Redesigning for new nodes costs Nvidia money, because most modern nodes no longer share the underlying technologies. Drake is halfway between Ampere and Ada. The smart money is that it will be on the same node as either Ampere or Ada.
Ampere was designed originally for TSMC 7nm. Nvidia spent a
shitload of money redesigning it for Samsung 8nm because of various problems (mostly cost/yield related) with TSMC 7nm.The only thing that made that redesign worth it was the transition from the data center versions of Ampere to the desktop version required some reworks anyway. Because of this, datacenter Ampere is still on TSMC 7nm.
Ada Lovelace was originally called "Ampere 2" internally, and is very similar to Ampere. The version of the Ampere that we
know is in Drake is halfway the version of Ampere in Orin, and Ada. Ada is on TSMC 4N.
The possibilities are:
- Drake is on 8nm, like desktop Ampere and Orin. Drake's GPU design was finalized after Lovelace's memory system was designed but before the tech was transitioned to 4N
- Drake is on 4N, like Ada Lovelace, and Drake's design was finalized after the transition to 4N.
- Drake in on 7nm, as an unusual halfway point. Nvidia reverse their redesign of desktop Ampere onto 8nm, pushing the tech back into the 7nm version, in order to give Nintendo a more power efficient version of Drake without the full cost of 4N.
- Nvidia ate the cost of a complete redesign of Ampere for Nintendo and Nintendo only - despite the fact that their desktop products sell 3-5x as many units per year as the Switch.
Options 1 and 2 probably cost roughly the same amount of money from a design perspective, but with different schedules. Option 3 is pricier than the first 2, and option 4 is pricier still. Which brings us to the second cost consideration, other than the node itself, which
@Hermii pointed out.
Nvidia keeps multiple products on the same node. It's not a law of Physics or anything, they can and will do whatever makes the most sense. But Nvidia has to buy the capacity, not the customer (which is usually also Nvidia, but in this case is Nintendo). Capacity is cheaper in advance, and
extremely expensive if purchase at the last minute.
By sharing nodes across products, if a product gets hot, Nvidia can back off production of other products temporarily in order to push more of the hot product out, without having to pay extortionist pricing. Conversely, if a product sinks, Nvidia's not stuck holding the back on all of the planned capacity they've already purchased, it can be shifted to other products. This is a second mark against an unusual node.
Nvidia is stuck paying for the success or failure of a
Nintendo product. That's a very unusual situation for Nvidia, and not one I think they'd be into.
But none of that is why I think 4N is correct. It's just because the Orin power consumption numbers roughly match with other power consumption tests on Ampere, and 8nm just breaks the power budget. Ada Lovelace levels of efficiency, however, hit right square in the middle of expected power draw.
I'm not an electrical engineer, so there are big error bars on my interpretation of the 8nm tests, and my sense of how much fat could be trimmed. And there aren't Nvidia GPU products on other nodes to test to see how close they could get, but since generally all the other likely nodes as less power efficient than 4N, I tend to lean one of two ways. "Either my understanding of power draw is right, and it needs to be on 4N to go as big as Drake, or it's wrong, and 8nm is viable. If 8nm is viable, then it's probably the biggest GPU-per-dollar that Nintendo could get, in which case, it's 8nm."