@boodream was asking some pages back about the implications of a foundry shift - TSMC to Samsung to Intel. I keep meaning to answer it, so gonna take a stab at it now.
If you understand the basic idea of what a node offers, you get the foundry differences. These three foundries offer different nodes, in different classes. Within a class of nodes one foundries product might be "better" than the other. If you don't, here is a very short summary:
Smaller process nodes make smaller transistors, which make chips smaller and more power efficient. Smaller chips mean a smaller device and/or you can pack more cores in there for more performance. More power efficient means better battery life and/or better clock speeds. Smaller and more power efficient also means less heat, which can indirectly make your device smaller, as it has to have less cooling. Smaller chips
used to also mean
cheaper chips, but that's not longer strictly speaking true.
There are lots of foundry companies, but there are only three that really matter in this discussion.
TSMC -
Taiwan Semiconductor Manufacturing Company, Limited - is the big dog, the largest foundry in the world, and they are "pure play". By "pure play" that means that they don't make chips for themselves - customers never have to worry about buying foundry capacity from a company that competes with them in another market.
SEC -
Samsung Electronics Co, Limited - is the second largest foundry in the world, and they are
not pure play, obviously. However, SEC has it's own corporate structure while being wholly owned, in order to maintain a certain amount of independence.
IFS -
Intel Foundry Services - has a similar model to SEC. They have a similar corporate model to SEC. This is Intel's second attempt at a foundry service, their first IFC, crashed and burned. There is an interesting technical/history lesson here about why these companies exist, but it's too big a tangent at the moment.
Practically speaking, IFS is off the table. Even if Nvidia were willing to put Drake over there (and there are lots of business and technical reasons they probably wouldn't), it's just not got a node advanced enough in a timeline that matters.
Historically speaking TSMC and SEC have gone back and forth on whose node in each class is dominant. But lately, SEC has slipped a
lot. This is not so much a quality issue as a timing issue. Everything in the 7nm class of nodes and beyond uses (partially or fully) a new kind of lithography tech called EUV. EUV machines were exceptionally rare at a time SEC's cash reserves were low, and TSMC just got more, and sooner. SEC invested a lot of effort in optimizing their older nodes, and cutting their cost, in order to make them competitive. That it worked
at all is kind of impressive, but it's left SEC behind on most fronts.
A short diversion on
node classes (or node
families). In the modern era, switching to the "next node" means not just improving the underlying technology, but changing some of it outright. Chip designs aren't compatible across the technology changes. As a result, foundries started making
half nodes and
sub nodes. Half nodes are big updates to node tech that doesn't change the underlying technology and doesn't require a design change (though you might want to anyway to get the most use out of it). Sub nodes are versions of nodes that have been optimized for a particular kind of customer.
With all that backstory, let's roughly rank the nodes by quality, and talk about the known differences between them. I'm not listing every node, just the most interesting for this discussion
10nm class: SEC 8nm
You start to see the confusion of naming
immediately. 8nm is the Samsung's most advanced half-node in the 10nm class. This is the least "advanced" node that anyone thinks is viable. Orin is on SEC 8nm, as are the whole line of RTX 30 cards. If you've seen power draw analysis on this forum about Drake, it comes from looking at Nvidia's other 8nm products.
Samsung is highly driven to offer deals, so despite being the largest chips, physically, it's probably the cheapest option. SEC doesnt list prices on their website, so we're guessing on that, but it's a good guess. And I don't mean "lowest price" I mean "best value." As in dollar per TFLOP, this is probably the best deal. If cost were the
only driver, this would almost definitely give you the most powerful machine, because Nintendo could pack in the cores all damn day.
7nm class: TSMC N7, SEC 7LPP
TSMC N7 is the node used by the original Xbox/PS5/Steam Deck chips. It's about a 30% improvement in GPU clock speeds at the same amount of electricity as SEC 8nm. SEC's 7LPP process is, as best anyone on the outside can tell, in the same ballpark at N7. However, it was late to the party, and N7 dominated.
As I said before - crossing node classes
or crossing foundries requires a chip redesign, and Orin/RTX 30 (Drake's sister chip and GPU parent, respectively) are on SEC 8nm.
However Ampere, the architecture behind RTX 30, was
originally designed for N7, and the data center products still use it. Nvidia redesigned Ampere for the SEC 8nm node for cost reasons, fairly last minute.
ARM, the CPU in Drake, has designs for multiple nodes, including these. So there is reason to believe that Nvidia might be able to cheaply get a design of Drake on N7. However there is also
7nm class: TSMC N6, SEC 6LPP
These are TSMC and SEC's half node upgrade over their initial 7nm products. They should be design compatible with their predecessor, and MS/Sony/Valve have all moved their consoles to N6. It's only slightly smaller, but it offers as much as a 30% improvement in electrical efficiency.
All three of these companies moved on mostly for cost reasons. TSMC has been retiring the original N7 node, but also the improved electrical efficiency allowed MS and Sony to lightly slim down their hardware, which was a cost savings for them both (lighter machines mean lower shipping costs, smaller heat sinks and fans means cheaper manufacturing cost).
I would call N6 the sleeper possibility, as it it an advanced, long-lived node that is cost effective, and is design compatible with N7, where Nvidia already has versions of Ampere and ARM designs.
5nm class: TSMC 4N, SEC 4LPP
TSMC 4N is a sub-node of TSMC N5. N5 is their first 5nm class node, 4N is a subnode that is exclusively for Nvidia's use. It's where Ada Lovelace is manufactured. It offers something like a 50% leap over N7 in both size and power draw, though it's a little hard to be sure, since Lovelace includes a bunch of its own internal power draw changes. SEC 4LPP is a Samsung 5nm class node that rolled out around the same time that 4N did, and I'm including it here mostly for completeness.
I said above that if price were the only issue, SEC 8nm is probably the best value. But it's not the only issue. We know how big Drake is, so we can estimate it's power draw based on Orin... and the numbers just don't look great! It's hard to believe an SEC 8nm product of that size can go into a handheld and offer anything resembling Switch battery life. But some benchmarks on Lovelace look almost exactly where you want them to be...
There is, of course, the cost of the redesign, but here we see that once again Nvidia has the tech on this process node. Lovelace was originally thought of internally as Ampere 2.0, and they are
very similar. And we know from leaks that Drake contains at least some of Lovelace's updates. It seems entirely possible that, just as PS5's GPU was essentially a "halfway point" between RDNA and RDNA2, that Drake's GPU is similar.
running smoothly at a decent resolution.
