So to my understanding we can estimate tensor performance of Drake using (the number of tensor cores per SM on Desktop Amere) * 4 *(the tensor performance per tensor core on Ampere) * 8 ?
So it could be that Nvidia designed Orin to have massive Tensor performance because of its nature as an autonomous car chip, and Nintendo thought the big tensor performance was not worth it and as such Drake uses the desktop Ampere design ?
Simpler, even. Sparse INT8 TOPS are 8x TFLOPS.
Yes we could! And the numbers are... very bad actually! But it's only 2 data points, and it's only for 4K output, so it's not super useful.
Yea, I'm assuming it's the same kind of tensor core. My question is, how many per-SM? Orin goes with a double rate tensor core, and 4/SM. In theory, if Nvidia/Nintendo wanted better FP16 performance (as it seems AMD/Microsoft went for) they could increase that number. Which would also improve DLSS times.
FYI, that isn't from NVN2. NVN2 is just a graphics API/driver.
Desktop Ampere, Orin, and Drake all have 4 tensor cores per SM. I can't speak to your performance findings, but both Nvidia's public documentation and the leaked source confirm the number for GA102 and GA10B, and the latter confirms it for GA10F.
They didn't do that with previous clean generation breaks, excepting where the market had already decided the old hardware was dead.
The part where Nintendo realizes they launched Switch way too cheap and missed out on billions of dollars. If they wish they'd launched for $350 back then, then $450 is probably the closest roundest equivalent for 2023 or 2024.
you have to balance that with happened to them with the 3DS. it's always more expensive to correct a mistake. The 3DS never really recovered.
I don't anticipate them letting go of the Switch name for some time. More likely to see some kind of superlative or descriptor like 'Switch XR'.
But they don't realize or wish that. The Switch is their most lucrative system ever. By the only definition that matters, they launched it at exactly the right price. This unknowable/counterfactual scenario would never be used as an excuse to push the price up when they have a proven winning strategy and no reason to fix what isn't broken.
3DS was a much bigger step, though. Even after inflation adjustment, 3DS's launch price was about 43% greater than DS's. If they were to pull that with Switch sometime this year, we'd be talking more like $550 than $450. So to complete the comparison I guess I'm saying 3DS might've gotten away with launching at $200.
Not cutting the price is part of the strategy. And they're going to repeat this strategy.
The PC part is definitely not true, GOW Ragnarok already got back its budget many times and sold 11 million in three months... So starting from there, what you're saying is nothing short than an overestimation of how much money Sony exclusives take to make. The truth is that many of their games weren't even more expensive than RDR2, which supposedly costed 100 million minus the marketing budget. Just keeping that clear for obvious reasons, only titles like TLOU2 have possibly costed all that much to develop and that's a big doubt, because Sony has shown to handle their budgets and dev times as good (if not better) as Nintendo. The brand and their exclusives have literally never been more profitable, that wouldn't be possible if they wasted as much money as you claim they do.
It's like people forget what happened with the 3DS launch all of a sudden. We have as close to an apples-to-apples comparison as one can get: new more technically capable hardware succeeding Nintendo's most successful hardware cycle to date, priced higher to try and profiteer off hardware day one. What was the result? OH YEAH, they slashed the price by over 30% within its first year at market, something they hadn't done since the Virtual Boy.
Edit: nvm, had the prices completely reversed. i somehow thought the OG DS launched $130 and Lite at $150, but had it reversed.
No, we're talking about a 33.3% increase instead, which still isn't small by any stretch of the imagination.
The "deal" on the TX1 comes from one of the earliest rumors of Nintendo going with Nvidia for project NX. IIRC, the source is known for editorializing, so it's of fairly dubious credibility.
Yeah, I feel like this framing is just too simplified to be of much value. You're compressing out the differences in the actual content and scope of the games entirely, which one of the big reasons BotW is such an outlier. BotW is a massive, heavily systems driven, open world game. Even with finished tools, it's just not an easy sort of game to make (see: TotK). Most Nintendo games aren't trying to be BotW and really don't need to be. They produce a wide spectrum of games, and are one of the most prolific publishers in the industry as a result.
I remember awhile back we were going over this in detail looking over these whitepapers and the only difference that could be found between Orin's Tensor cores and GA102, was that increase in cache per GPC. Both utilize 3rd generation Tensor cores and nothing else made much sense, other than the increased cache allowed the GPU to perform better.
Except the DS market had already been taken away by smartphones and the 3ds released with an unappealing gimmick that was said could hurt children's eyesight. This hasn't happened with the Switch and the Switch OLED showed the exact opposite: there is a large demographic interested in spending more for a better hardware.
If the market was taken away by smartphones or because of stereoscopic 3D, no price decrease could have created the seemingly rapid turnaround in sales that occurred. The market shrank overall, yes, that is undeniable, but the price for those potential consumers that remained was still repellent, considering it could only sell 710K units in its first full post-launch quarter worldwide. The same quarter the following year, sales were up 163% YoY, more than 2.5x more. Consumer attrition to smartphones can only explain an overall total lifetime sales decrease, not why consumers weren't buying it until a price drop.
3DS sales started to turn around when Nintendo not only cut the price, but also started releasing desirable software rapidly.
40% of all OLEDs sold are sold to people who already own a Switch. They may not be Nintendo enthusiasts like us, but that figure certainly indicates they're Switch enthusiasts, especially when 2nd Switch purchases across the entire lineup account for 25%. When you account for the fact that 30% of Lite buyers are buying their 2nd device, in terms of attracting new buyers, the non-OLED non-Lite model appears to still rule the roost worldwide.
If the games will only "take better advantage of the hardware as now on" with the massive amount of time and apparent effort they already take to make (which is already a lot more time than other exclusive releases, funnily enough)... You will be kind of giving me the reason about them being unable to make comparable console games anytime soon even with abnormally long dev times and up-to-date systems compared to the rest, simply because (plot twist) they're less experienced at making these kinds of games. Even then, it's been already proven many times before graphics are only a small part of why the AAA space is supposedly unsustainable (which is not even true for the studios that make them), it's the stupidly massive amount of content they usually have to extend their runtime as much as possible, dwarfing that of games from 15 years ago. That has always been a poor argument from someone that didn't know a single thing about how the development pipelines actually work. Not a good sign for the thread you're using it in its most simplistic form and people even vote you for it.
From your own article: "which is ten percent higher than the amount of Switch Lite repeat buyers". 30% of Switch Lite buyers already own a Switch.
Orin has wider tensor cores, it's similar to GA100. It's all to do with the theoretical performance rather than efficiency gains or whatever, although it sounds like the larger shared memory is necessary for that.
I will preface it that I agree with the main point that Nintendo won't have big profit margins at launch. If Reggie interview is to be believed, the 3DS cost them roughly $220 at launch and NCL was fine selling at that price but not at loss ($199). So, there was never really a time where they were too greed with launch profit margins. And with how much they make from software, the gains are not worth the risk.
Man, you are REALLY overestimating the time it takes to make graphic assets in the production of a game.
Oldpuck can clarify about the performance testing, but the number of tensor cores per SM should be the same for all Ampere GPUs at 4, and what (we think) we know for sure about Orin's tensor cores is that they have the same performance as desktop Ampere (and Drake) except for various half-precision operations, where Orin has exactly 2x the performance.
Indeed, that's what the specsheet indicates.
It’s not an either-or situation. The neural network is trained, and some of the weights will be near zero. Sparsity acceleration clips up to half of those near-zero values by skipping the multiplication. You can’t predict a priori how those sparse weights will be distributed in the network.
Thanks for the explanation. So, if I understand correctly, the best option to estimate DLSS cost on Drake would be either to find a point of comparison with Ampere (I don't know if we have) or find some Nvidia paper where they say "Sparsity is so cool and awesome it gave a X% improvement to DLSS".
I'm not sure you're even reading what I'm saying, because you're responding to a post where I literally point out that BotW's content is not particularly representative of Nintendo's output as a whole.
I double checked this, and I realized I was actually wrong on one of the significant points. Sparsity acceleration does actually always skip 2 out of every set of 4 neighboring weights. The accuracy of the neural network predictions will be significantly affected if you prune weights that are not close to zero, but it will have a 2x throughput increase no matter what. Sorry for being misleading!
Oh okay. So the idea is the sparsity algorithm starts with a dense matrix, and doesn't just cut values near 0 but modifies the matrix such as it's possible to cut 50% without modifying too much the output. But obviously there's still some loss.
I am completely lost. Is it 2x or up to 2x ? Is it supposed to mean that desktop Ampere supports 2:4 sparsity, but some other Ampere GPUs don't support that but still support sparsity in some manner ? Is the second part meant to simply do a comparison with Turing and isn't supposed to be related to the first ? WHY DOES NOTHING MAKE SENSE ?
There were reports of some sort of "cancellation" of a device that may or may not have contained Drake, but the details don't make a whole lot of sense and there seems to be quite a lot of speculation mixed in. I'm mostly ignoring it for now, since the information seems too garbled to draw conclusions from and seems likely to be tainted by bad assumptions.
There's no real change from what you've previously read. Release dates are up in the air but it's a lot of "I thinks" than any concrete info
