-
Ever wanted an RSS feed of all your favorite gaming news sites? Go check out our new Gaming Headlines feed! Read more about it here.
-
We have made minor adjustments to how the search bar works on ResetEra. You can read about the changes here.
New Switch model coming in second half of 2019, reports Japanese WSJ
- Thread starter Crushed
- Start date
- Status
The Tensor cores take up too much room, I forgot the die size for Titan Volta was so big, my point still stands that Nintendo would be better served with more cuda cores than tensor cores. For instance the GTX Titan X(p) has 3840 Cuda cores and 471mm^2 die size, meanwhile the RTX 2080 TI has 4352 Cuda cores at 754mm^2, Tensor cores were a huge difference in die space, and they are useless for running games in a traditional sense, they basically act as REALLY good AA. That isn't worth the space for more cores here, you could have a Tegra die with ~140mm^2 that has 512 Cuda cores, adding Tensor cores would increase that size to ~175-180mm^2 and that is just to have ~1/10th of the available cores in a RTX 2080TI. Would DLSS even work well with so few cores? that is a lot to ask when a chip that size could probably house 1024 cuda cores. Remember the GTX 1060 is only 200mm^2.
I do think you are probably right Thraktor, it's what I've speculated all this time and as for Skittzo0413, I believe it's possible that a dev wouldn't know they have a super powered devkit, and even if they did, that really isn't out of the ordinary for devkits to have a piece of hardware that mimics the performance but is much bigger, like the Wii U devkits using R700 HD 4800 series GPUs. It's also probably worth noting that devkits can't be opened by a developer, so whatever hardware is inside a devkit is only known through whatever specs Nintendo sends them, and Nintendo could very well disable "x" SMs on a GPU and list the devkit with whatever is left available to them.
So no to Tensor cores being likely IMO, one last nail in it's coffin is that it uses power to have tensor cores there, and when they are used, it would be a much higher power draw than with them not there, and since the biggest constraint is the heat and power of the device, Tensor cores as a performance solution make very little logic over more cuda cores.
My point here and the point I've made in the past is that in the same space that Xavier has allocated to GPU (90 mm^2 at 12nm) the Switch Pro could fit a Turing GPU with 512 CUDA cores and their corresponding Tensor Cores and RT cores. Of course, they could have more CUDA cores in the same space if they only had traditional rasterization, or they could have the same CUDA cores, but at higher clocks. In this line of thinking, perhaps it'd be better to have 512 CUDA cores in a 7nm node for a small, cheap chip that can clock very high than to have a big 12nm chip for the same price with all the Tensor cores and RT cores.
If DLSS scales well with resolution, then maybe the tensor cores could be worth it, since Ray tracing does scale linearly with resolution. But there may be other factors to this.
A big one is actually, how mature and widespread is Nvidia's RT implementation going to be? Because so far Playstation and XBOX and most PCs won't have it, it's locked to Direct X and to Nvidia's proprietary software, etc.
Just like Nvidia and AMD pioneered different forms of GPU compute and it took quite a while until it was more that a footnote, we might have a similar situation here. Look at Physx. It lives on through GPU compute and now everyone uses that, but it took an independent and vendor-agnostic solution.
It's an additional functional unit built into the SMs (shader modules) alongside the regular logic. If you didn't explicitly call on the tensor core to do something you wouldn't know it was there.
It's possible that the vast majority of them were for use inside Nintendo. If they were just looking for feedback on possible hardware configurations then they may only send out a handful of them to key third parties, which means less likelihood of leaks.
I disagree with the bolded, I think there are plenty of unexplored uses for neural-net accelerators in games, particularly in visual processing. An obvious one is the use of style-transfer or similar techniques for non-realistic rendering in games, which is something which I think could be particularly appealing to Nintendo. I'm just not sure it's something Nintendo will be willing to go for just yet given their die size and power constraints.
Keep in mind that the entire Tegra X1 die size is 121mm2. Even if they could cut down the remainder of Xavier's 350mm2 by swapping the Carmel cores out for A53/A73s, dropping the automotive-specific hardware, etc., they'd still likely be on the north side of 200mm2 on a 12nm process, which is a big die for a device like the Switch. At 7nm it becomes more reasonable (TSMC claim 70% die size reduction compared to 16/12nm processes, but I'm always kind of skeptical of those claims), but even a ~120mm2 7nm die wouldn't be cheap in 2019, it would place it alongside the top of the line mobile SoCs from Apple, Qualcomm, etc.
Nvidia's PC ray-tracing uses the DirectX Ray-Tracing API, as far as I'm aware. In any case, I just don't see RT hardware hitting the performance levels necessary to be useful in a Switch-like device any time soon, as much as I'd love to see consoles with dedicated ray-tracing hardware.
That's what I say. The way I see it they can have a "wide and slow" chip at 12nm with the tensor cores and the RT cores and so forth, for ~200mm ^2 (Which by late 2019 will be cheap as hell) or a "narrow and fast" chip without Tensor Cores or RT cores, but the same 512 CUDA cores, in 7nm and running higher clocks.
One thing I don't see people talking about are changes to the joycon. Do you think we could see the shoulder buttons changed? Remember that one patent with the scrolling clickable wheels as a shoulder button? Maybe that could be a thing.
If you've seen it in a patent it's pretty unlikely they're going to use it in the future. If they were they wouldn't want to publish it beforehand.
I see. It looked so cool. I would love to see evolutions like that.
It wouldn't be unprecedented with Wii Motion plus and Circle Pad Pro both coming out in less than 3 years, buuut I can't say I'm expecting anything this time around.
There's a reason you don't see phone and tablet SoCs much bigger than 100mm2 or so, and it's not just the cost. Die size is directly proportional to static leakage, which means increased idle power consumption, and a reduced benefit from lower clock speeds. Even at 12nm, I'm not sure if they'd be able to clock a 200mm2 chip low enough to fit within the ~5W TDP they'd be looking for in portable mode, which would mean a battery life hit compared to the current model.
Docked would be a different matter, and if Nintendo wanted to release a stand-alone home console then it'd give them a nice affordable competitor to PS4 and XBO, but I think a 4x SM configuration is just too big on 12nm to work in a hybrid form-factor.
I'm sure it was prototyped quite a bit and something similar might pop up in the future but there's clearly a reason they didn't go with it. Maybe it was just uncomfortable to actually click or took up too much room internally.
Still waiting on that official Joycon-L with a dpad they pattented over a year back. :/
Hori's design is slightly different. No gyro, wireless or SR/SL. Also not a Nintendo dpad.
I thought yes, because Hori sold the D-pad joycon under the license with Nintendo, right?
I see.
So, unless Nintendo goes directly to 7nm, we're fucked in terms if hoping for Xavier like specs? So it's either 16nm X2, or 7nm for a chance at 512 cuda cores?
not really. licensed in that Nintendo officially allows them to use their stamp of approval. but licensing a patent is different.
If they aren't suing Hori for infringement then yes they have effectively licensed the patent.
They are not. They aren't even allowed to call them Joycon. It is simply called a D-pad controller.
They did acquire the Mario/Zelda license, but not the joycon license. Instead it's still just a 3rd party accessory.
Though Nintendo did include a special joycon update for Horizontal to fix the battery drain issue
or they aren't using the patented design, as (almost) all third party controllers do.
AFAIK they didn't file a design patent for a joycon with a D-pad, they just had it in a utility patent. The design is completely immaterial in this case, the only thing that matters for an infringement case would be whether or not the Hori controller could read on the claims of Nintendo's patent.
I don't recall off the top of my head if that joycon was described in a claim or just a drawing of one embodiment in a general Switch related patent but if there was any claim drawn to a detachable controller of any kind then any third party controller would indeed read on that and technically Nintendo could sue. Them not doing so is usually a sign that they licensed the patent.
Or the claims all include the Switch base device along with the detachable controllers so single joycons wouldn't read on them. I don't have the patent(s) in front of me.
I'm going into the hospital in a few, my wife is due with our second child, so I probably won't be keeping up with this thread very much for the next week or so.
Couple problems here IMO. Tegra X1's cuda cores are in 2 blocks of 128 clusters, size is only 6mm^2 a piece, yes there is other components to the GPU, but Xavier is not a good indicator of such a GPU size as Volta's GPU in Titan V is 815mm^2 with 5120 cuda cores, well half (2560) the cuda cores with GTX 1080's die is only 314mm^2. GTX 1060 is 200mm^2 with 1280 Cuda cores, if we were to just linearly look at cuda core to die size, 1280 cuda cores + other components took up an extra 114mm^2, that would reflect less than 24mm^2 for another 256 cuda cores for 16nm over Tegra X1's 256 cores. It seems 12nm without size reduction (which TSMC claims there is some size reduction) you'd only have ~144mm^2. As for leakage, 20nm is far worse than 3D transistor solution. I don't see how a reasonable estimation for a Maxwell/Pascal GPU solution SoC with even the large A57 cores (A73 cores are what? 30% as big) could ever hit 200mm^2. We should be looking at ~23mm^2 for the extra cuda cores at most.
Dude, you dont need to say 'excuse me" in a internet forum, we are just speculating anyway. Congrats in your child.
Thanks, I'm just saying that I might not get back to this thread to help with my estimations in the future, but yeah it is a bit ridiculous, though I've known some of the people in this thread for 7+ years now.
Well, the die sizes I'm talking about there are including the tensor cores. If we exclude these then the die would be smaller, but I still think 4 SMs is probably a stretch on 12nm. As I mentioned before, 3 SMs with a modest clock boost would probably be plausible, and hit 2x the performance of the existing Switch, so not exactly a bad return for a new model only two years after the original.
Congrats! Don't worry too much about keeping up with Resetera, there are more important things in life than speculating on upcoming Nintendo hardware.
Yes, Xavier isn't the best example, but the question was about a GPU with tensor cores I believe, and Xavier is the most readily available SoC to compare against. I still think 4 SMs is still a stretch even without the tensor cores, though. My guess is that Nintendo will try to stay as close to 100mm2 as possible, both for cost and power consumption reasons.
Also, the claim of a 70% die area reduction for A73 over A57 is based on manufacturing process going from 20nm to 10nm. On the same process, the die area drops about 10% from A57 to A72 and then "up to" 25% from A72 to A73. These are also core-only, so for a full core complex with L2 cache you might be looking at a 20% or so reduction in size going from A57 to A73. Which is still pretty good given A73 is both more powerful and more power-efficient.
The other possibility which we haven't talked about is that Nvidia pushes Nintendo into using their new Carmel cores. Judging by the Xavier die photo, though, they're pretty huge, and likely designed for higher power-consumption than Nintendo would be looking for in a Switch-like device, given Xavier's TDP.
I think with the Switch+ we'll get Docked Mode resolution/graphics in portable mode (downsampled to the 720p screen) with better Performance and while docked better Performance overall and games with dynamic res like Zelda, DOOM,... stay at the highest res.
I play handheld only, so can't vouch for what I'm about to say, but I've read more than a few times for various games that their handheld version runs better than their docked version (so, the developers optimized performance for handheld, not for docked).
That apparently being the case, if Switch+ gets docked performance, I wonder if could we end up with a monkey's paw situation where some existing games actually run worse in Switch+ handheld mode...
I am not familiar with what Liam said. Was that pure speculation or speculation based on info/hints he has?
I'm basing this off people transcribing his paid podcast (ironic, no?) in another thread. he seems to characterize it's more like the DS Lite, a form change rather than a spec bump like the DSi or n3DS
Ugh, not sure if I would get an "upgrade" as minimal as that one. Only some sort of Switch Pro would make me bite the bullet.
This is where Nintendo shines. Never expect the unexpected.
Well, if it's indeed more of a form change, then regardless of power level, that sounds to me like the 3DS successor I expect to come in time for "Classic" Pokemon and especially Animal Crossing - a lower cost handheld that plays Switch games.
If that will happen, while personally I'm fine with my Switch, I will welcome it as the third and fourth Switch in my household - my two girls that don't have a Switch yet really don't need a $300 Switch that can dock, there are already 2 of these in our household. Both just need a handheld machine that can play Switch games, hopefully even more drop proof than the current version. And, if indeed Nintendo is coming out with such a version, that is exactly what Nintendo hope to get from it (that plus 3DS owners who hold out due to Switch price point/form factor).
One thing I noticed is that both Matt and John said that such a version exists, not when it is coming, or even if it is coming. For all we know, Nintendo could have been playing around with a pro version from the onset (including giving it to a select few major developers to play around) and then decided not to come out with it commercially, and this would not be a contradiction to what these insiders said. They only said "yes it exists". Or there could be two versions coming, "lite" in spring 2019 and "pro" holiday 2019 or some other date, and again this would not be a contradiction to what these insiders said.
Ordinarily I'd expect a DS Lite like revision from Nintendo at around this point, but there's not much Nintendo can do in that department while retaining compatibility with the existing joycons (and I very much doubt they're going to break compatibility with them after just two years). I can see a smaller portable-only device as a definite possibility, but it would be a separate device alongside the Switch, not a revision of the current model.
insiders never said what the revision is, just that there is one
in either case, a Lite and Pro would be based on the same SoC, so one would naturally beget the other
aye, this is the only way I can see a Lite model happening. even with a shrink to 12nm, there's not much that could be changed, size-wise unless you make a model with a smaller battery, but the capacity would have to change a lot in order to have a non-negligible shrink
Liam Robertson is a terrible "insider". He's been hilariously wrong ( that last minute Star Fix Zero delay, anyone ?) over the years.
Ill buy twoWant N to take inspiration from this and call it....Micro Switch
I could see a Lite style revision maintaining size and shape while improving other ways internally and externally from two things: improving the screen and moving to a smaller fab. With those a Lite unit could be lighter, brighter, slightly increase screen size/reduce bezels, move to passive cooling and increase battery life.
I tend to think we're getting some level of performance bump too though based on what insiders here are saying. These are all improvements that could exist alongside that too as well as other things like increasing storage. Nintendo will want to have enough value to maintain a $299.99 pricepoint as PS4 Pro and Xbox One X drop to that price and their successors launch.
- Status