I think these specs especially the storage speed, 8gb, and very mid gpu is a worse case situation and I’m just expecting 12gb minimum because 8gb really isn’t a lot. 16gb of unified lpddr5 or 5x with a super lightweight os would be ideal.
That’s very pesmistic thought about clocks and even ram beacuse if I understand correctly you expect just 8GB RAM lol, I don’t think Matrix Demo or BOTW 4K/60FPS would be possible on this specs even with DLSS
Yeah also Nintendo knows next gen games are going to take more time, money, and people and they need 3rd party aaa games to fill in the many months gaps between release so it’s in there best interest to make a somewhat powerful device.
Samsung’s newer process nodes are know to be super inefficient for portable SOCs.
Samsung foundry is shit and their 5nm is much worse than TSMC 5nm
Samsung 5nm doesn’t make any sense, not just for the reasons listed above, but because Nvidia doesn’t make anything in it. It’ll either be Samsung 8nm used in Ampere chips or TSMC 4nm used in Ada Lovelace chips.
Yup.
Btw SEC8N is just optimized 10nm in Samsung terms xD
Unless every single manufacturer colludes, at least tacitly, to keep the prices high, it seems pretty unlikely that the current prices could hold with the introduction of a new customer base that is both far larger than the current one and much more price sensitive. Being used in a device like a Switch could even be the tipping point that leads to mass adoption of the format.
Again, Erista models from ~June 2018 onward aren't vulnerable to the USB exploit in RCM mode. New Switches weren't vulnerable to soft mods for over a year before Mariko officially released in a device; hacking was not a driving factor in getting the 16nm models out.
Am I really living in a world where 3 hours of battery life is lopsidedly underclocked? It was the most they could get away with at the time, and that process shrink was years away. I wouldn't expect the new model to have less battery life than the launch Switch.
Around the time that the 40 series cards started development. I don't know when that would've been, but we do know that T239's design was finalized at the same time as the 40 series, with tapeout occurring in mid-2022. It's not at all hard to imagine that 4N was an option to Nintendo when development started.
Chip design costs are mainly a function of the number of transistors and of the node used. The physical chip area is not a direct factor, unlike with manufacturing costs.
Precisely, I was replying to @Concernt 's post about 4N being cheaper out of the gate. If IC designers and their customers only cared about the raw cost of manufacturing a single chip, nobody would ever design a chip for anything other than the very latest and most efficient process node, because all else being equal, a given design will always be cheaper to manufacture on a newer process. But all else is rarely equal: the design cost, photomasks and prototyping for a cutting-edge node can be considerable, and in our case Nvidia also paid TSMC a hefty sum simply to reserve all that 4N capacity, some of which would presumably be passed on to Nintendo. So yeah, using 4N is not some kind of free lunch, but at the scale we're talking about here, it probably makes sense.
Nvidia's custom 4N is just TSMC's usual service to clients. The fab process is tuned to the product so no two 5nm products are exactly alike. AMD's 5nm products also have special changes. Nvidia applies marketing to this process to make their shit sound special
This is true. But wasn't the rumor that Drake was taped out around the same time that the Ada Lovelace GPU was? If that is true, Nintendo could have opted for Ada GPU but from the Nvidia leak, Drake has some power saving features from Ada.
What Ada GPU? There's no SoCs on Ada yet, and Thor is likely still a few years out, on top of the R&D to turn it into a low-power gaming chip. For Nintendo's purposes of a tablet with a 2024 release, an Ampere SoC based on Orin was the best that Nvidia could offer.
In short, the benefits for a handheld from Ada are mainly in the realm of efficiency. Since it inherits those from Ada, along with, it seems, possibly a more efficient node, there's no real reason to worry about whether T239 is Ampere or Ada. In reality, we already know it's neither desktop Ampere nor desktop Ada, it's a version between them. This sounds bad, but it really isn't, as the additional features of Ada aren't relevant in the market segment and performance target of T239.
And the sheer size of the chip. Before the Nvidia leak, we were all convinced it would be 8nm. We speculated it would be between 4 and 8 SM, with 8 being a long shot.
Don't you get more chips per waffer ?
Older nodes have higher yields
Isnt SEC8N yields notoriously bad?
Which I mentioned already and you're conveniently glossing over. 2.2x more expensive per wafer, but 2.7x more denser for TSMC 4N.
Any of them. Drake was found to be tested by folks who tested Lovelace gpus on LinkedIn
Not all the time and not in this case. Samsung 8 nm in particular is quite infamous for its low yields compared to the equivalent TSMC 7 nm node. Meanwhile, TSMC 4N is an exceptionally high-yield node for its type at around 80%, compared to 75% for Samsung's 4 nm processes. While that may not seem like a lot, when you're talking on the order of tens of millions of chips, that small difference adds up fast.
Isn't that similar to the logic used for the Switch 1 being on 16nm, only for it to be on 20nm anyway?
No, 20nm was just abandoned. But Nvidia already ordered wafers to be used. The problem here is, does Nvidia still have wafers to be used on Samsung 8nm? Because we know they do for tsmc 5nm
That's not really comparable, since Nintendo needed something quick, and there were already X1 chips in 20 nm available. Nintendo can afford to be a lot more choosy about node now.
You can't just stick a graphics card in it and call it a day. The whole point of an SoC is that you have the CPU and GPU all on one piece of silicon.
Yes. SEC8nm is a dead end which already has been squeezed every optimization it could get. Further shrinks to SEC/SF 7/5/4 would require a design port, new verification and validation, etc. It's the usual stuff that is done when you shrink a design. It's just that it became more and more expensive with modern nodes. The biggest question a lot of people ask is: Does it make sense for Nintendo to manufacture a big SoC on 8N to "save money", just for them to redesign it again and shrink later and waste more money? Why not go with 4N from start and stay there until Switch 3?
TSMC 4N has better parametric yields and similar catastrophic yields as of 2022. It's even better now.
I think what Feet is saying is that the people who tested the Ada Lovelace GPUs also tested T239, giving credence that T239 was taped out around the same time and that T239 might be on the TSMC 4N node that the Ada Lovelace cards are, even though T239's GPU is not an Ada Lovelace GPU.
Are there any concrete statistics on node yields out there? That'd be very helpful in future node debates.
So Nintendo was in a bigger rush then compared to now.
if Nintendo was in a similar rush like before the Switch launched, then they just might use Orin but with parts fused off
I believe that this is not the type of patent that we can discard according to the pattern "if it was made public then it will not be used", this time it is not a ginmick, a new form-factor or anything like that, this going public before the announcement in no way gives a spoiler of what's to come. I don't think anyone is surprised that Nintendo cares about more efficient forms of cooling, even if the Switch 2 doesn't necessarily need it.20W Switch 2 is real @Concernt
In all seriousness, since this got patented, they obviously didn't go with it. So what have they chosen instead?
