Support NeoGAF

[KingSnake]

Pascal isn't ready
Then why not producing Maxwell chips at 16nm, it's the same shit anyway. Okay.

I think you lost track of your line of thought. Or you don't see the quotes.
Why are you jumping now back to Pascal when you were talking about Maxwell in the post that I quoted?

And what that has to do with the actual thing I quoted? About Nintendo beating Nvidia at their own game. How is Nintendo beating Nvidia when Nvidia are still the ones doing the design? That would be Nvidia beating Nvidia.

Why are you bringing the derogatory tone in your posts whenever you don't have a proper answer? Do you think that it goes unnoticed that you don't have a proper answer if you try to be patronising? It doesn't.

 

[blu]

If Nintendo is forced to go with Maxwell because Nvidia isn't ready to provide Pascal chips then they aren't going to get a Maxwell with all the keyfeatures of the Pascal series either.

Can you list those keyfeatures for me?

 

[Hermii]

One thing I don't get that if the specs in the op are so accurate and the specs in the op is a stock tx1 then what's up with it supposed to be a heavily customized chip?

 

[KingSnake]

Can X1 be clocked high enough to do 1TF at fp32 or is that likely impossible? So either way we are looking at a docked 1Ghz X1 with 512Gflops and portable mode 256Gflops?

Theoretically it can get more CUDA cores and a higher clock. Practically it probably won't happen because of battery and heat and the author of the article is mixing FP16 and FP32.

 

[PdotMichael]

I think you lost track of your line of thought. Or you don't see the quotes.
Why are you jumping now back to Pascal when you were talking about Maxwell in the post that I quoted?

And what that has to do with the actual thing I quoted? About Nintendo beating Nvidia at their own game. How is Nintendo beating Nvidia when Nvidia are still the ones doing the design? That would be Nvidia beating Nvidia.

Why are you bringing the derogatory tone in your posts whenever you don't have a proper answer? Do you think that it goes unnoticed that you don't have a proper answer if you try to be patronising? It doesn't.

Nintendo gets a Maxwell chip with a better power efficiency than all the other companies who get the newer Pascal chip. That is pretty silly.
Also you are the guy who just deleted my entire post when you quoted me.

With Switch now a >1TF machine and Nvidia flops are better than ATI flops. I guess we are again at the "better than Xbox One" point.

 

[KingSnake]

Nintendo gets a Maxwell chip with a better power efficiency than all the other companies who get the newer Pascal chip. That is pretty silly.
Also you are the guy who just deleted my entire post when you quoted me.

I deleted it because I wanted to point out how ridiculous is that particular statement.

Who said anything about a Maxwell chip with a better power efficiency?

 

[PdotMichael]

I deleted it because I wanted to point out how ridiculous is that particular statement.

Who said anything about a Maxwell chip with a better power efficiency?

It's what you need if you want to hit that performance goal.

 

[KingSnake]

It's what you need if you want to hit that performance goal.

No. As I stated already several times, at least once in an answer to you all you need is more CUDA cores and/or a higher clock. That's not power efficiency. In a purely theoretic scenario it could have that while plugged in and running at a much lower clock on battery. That's purely theoretic.

I also stated that I don't think that's the case and that >1TF is most likely FP16.

 

[PdotMichael]

No. As I stated already several times, at least once in an answer to you all you need is more CUDA cores and/or a higher clock. That's not power efficiency. In a purely theoretic scenario it could have that while plugged in and running at a much lower clock on battery. That's purely theoretic.

I also stated that I don't think that's the case and that >1TF is most likely FP16.

You are still stuck with the same form factor that means the Maxwell chips needs to get a higher performance per watt than what Nvidia presented for Pascal.

And if you don't believe >1TF than what is your point of arguing?

 

[blu]

You are still stuck with the same form factor that means the Maxwell chips needs to get a higher performance per watt than what Nvidia presented for Pascal.

And if you don't believe >1TF than what is your point of arguing?

Ahem, ping?

 

[GhostTrick]

If Nintendo is forced to go with Maxwell because Nvidia isn't ready to provide Pascal chips then they aren't going to get a Maxwell with all the keyfeatures of the Pascal series either.

There isn't that much too talk. And even if they would get. Getting Maxwell to provide "more than 1 teraflop" in that formfactor would be quite the task. But Nintendo is gonna beat Nvidia in their own business.

Yes, Maxwell can provide 1tflop. And Pascal can too. People in this discussion makes a silly shortcut as "X1 = Maxwell, X2 = Pascal".
Nintendo hitting 1tflop (ain't happening btw) isn't relevant to the choice of Maxwell or Pascal but the choice of 2 or 4 SMM. 4SMM Maxwell will outperform 2SMM Pascal, at the same clockspeed.

But because of heat constraint, it's highly unlikely Nintendo takes over 2SMM. With 2SMM, whenever it is Maxwell or Pascal, it's still 2SMM. It's never going to be a 1Tflop part with so few shader units.

Basically, Nintendo going Maxwell instead of Pascal would mean either 25% less faster device... or less battery life.

 

[KingSnake]

And if you don't believe >1TF than what is your point of arguing?

You are the one obsessed with the 1TF. I was just answered to specific points in your posts which were wrong or ridiculous and you keep coming back and talk about the Nintendo fans and 1TF. Even quoting just the respective parts from your posts I was specifically addressing didn't work. "But, but the Nintendo fans fantasising about 1 TF!"

 

[Nightbird]

So what Thread was Emily talking about that was 90% correct?

This one?

 

[PdotMichael]

Yes, Maxwell can provide 1tflop. And Pascal can too. People in this discussion makes a silly shortcut as "X1 = Maxwell, X2 = Pascal".
Nintendo hitting 1tflop (ain't happening btw) isn't relevant to the choice of Maxwell or Pascal but the choice of 2 or 4 SMM. 4SMM Maxwell will outperform 2SMM Pascal, at the same clockspeed.

But because of heat constraint, it's highly unlikely Nintendo takes over 2SMM. With 2SMM, whenever it is Maxwell or Pascal, it's still 2SMM. It's never going to be a 1Tflop part with so few shader units.

Basically, Nintendo going Maxwell instead of Pascal would mean either 25% less faster device... or less battery life.

It doesn't matter what is somehow possible.

We know the form factor, we know the basics of Tegra Maxwell and Pascal, we also aware of the price range a dedicated Nintendo handheld is supposed to operate

1TF is out of range - good that we all share the same opinion in that regard. Now people can discuss what is more likely. 512gflops witha a 128bit bus for the ram would be like a best case scenario right now.

You are the one obsessed with the 1TF. I was just answered to specific points in your posts which were wrong or ridiculous and you keep coming back and talk about the Nintendo fans and 1TF. Even quoting just the respective parts from your posts I was specifically addressing didn't work. "But, but the Nintendo fans fantasising about 1 TF!"

So you are full of nonsense. Okay.

 

[Hermii]

So what Thread was Emily talking about that was 90% correct?

This one?

Yea

 

[Roo]

Ahem, ping?

Gotta love how he keeps ignoring this post lol
Them Pascal key features

 

[tkscz]

One article, written badly with no history of being correct on rumors, brings up the specs of what we pretty much saw as the devkits and not final product, and everyone believes him?

 

[Herb Alpert]

One article, written badly with no history of being correct on rumors, brings up the specs of what we pretty much saw as the devkits and not final product, and everyone believes him?

Of course, it's bad news (sort of)

 

[Donnie]

Heavily modified Maxwell?, erm correct me if I'm wrong but Pascal is a very slightly modified Maxwell (its basically a Maxwell on a smaller process).

What exactly are our options with bandwidth if its maxwell now?

The options are unchanged, they both support exactly the same kinds of RAM on the same kinds of bus.

 

[Donnie]

If Nintendo is forced to go with Maxwell because Nvidia isn't ready to provide Pascal chips then they aren't going to get a Maxwell with all the keyfeatures of the Pascal series either.

What key features are you referring to?, please list them.

 

[PdotMichael]

What key features are you referring to?, please list them.

Like 16nm, higher memory bandwidth, new color compression, Denver II CPU and stuff. Google is your friend.

 

[Donnie]

No way they would make 16nm maxwell LOL... They might as well just use Pascal at that point.

Nintendo/Nvidia have been customising this SoC for at least two years, way before Pascal ever happened. Pascal is basically a 16nm Maxwell. Sounds like this chip is something different.

 

[Trike]

How many friezas is this switch

 

[Donnie]

Like 16nm, higher memory bandwidth, new color compression, Denver II CPU and stuff. Google is your friend.

They're all "features" of Parker, not Pascal. Thinking isn't your friend.

 

[Metal B]

How many friezas is this switch

The Switch is Son-Goku, which can go Super Saiyan when docked.

 

[DaSorcerer7]

They're all "features" of Parker, not Pascal. Thinking isn't your friend.

 

[Donnie]

Haha, yeah well nothing irritates me more than someone who says something stupid and then acts as if you're the one being stupid "google is your friend" ffs

 

[MuchoMalo]

As someone who isn't the most knowledgeable about hardware, is it a huge deal if the Switch uses Maxwell instead of Pascal?

The only potential loss is in battery life compared to reasonable expectations. All of the people going "WII U PRO CONFIRMED!" have no idea what they're talking about and you shouldn't pay attention to them.

One article, written badly with no history of being correct on rumors, brings up the specs of what we pretty much saw as the devkits and not final product, and everyone believes him?

Emily vouched for it, or at least the Maxwell part. Whoever wrote it clearly has no idea what they're talking about and thinks that writing a lot makes them seem smart, but that doesn't say anything for their sources.

The long and short of it is, we should expect the specs in the OP of this thread to be 100% correct, and we should expect no more than 3 hours of battery life for the Switch. If we get better than that, yay.

 

[Cuningas de Häme]

Reminder here too, that article speaks about Maxwell AND 1TF power, while in here people have been discussing 750GF (at most!) and Pascal.

So, which ever is true, it's a win for us.

And please someone put that poor fellow out of his misery already!

EDIT. Yeah, my flips beats everyone else's flops... I'll just go lie down for a moment.

 

[disap.ed]

Nintendo/Nvidia have been customising this SoC for at least two years, way before Pascal ever happened. Pascal is basically a 16nm Maxwell. Sounds like this chip is something different.

Pascal also didn't happen over night, these things take time for whomever they may be used.

 

[ggx2ac]

Reminder here too, that article speaks about Maxwell AND 1Tf power, while in here people have been discussing 750Tf (at most!) and Pascal.

So, which ever is true, it's a win for us.

And please someone put that poor fellow out of his misery already!

I think having the GPU perform 750 TFLOPS would be a bigger win than 1 TFLOPS.

 

[Skittzo0413]

That article sure did manage to ruffle a lot of feathers without saying anything at all.

It claims that they couldn't use Pascal due to timing issues, but does that make any sense to anyone?

Also, as far as 16nm Maxwell is concerned, 16nm processes have been available for a long time, right? At least most of this year? 16nmFF+ or 16nmFFC might not have been available as long, but in general, for a product being mass produced in late 2016/early 2017 what possible reason could there be to go 20nm? 16nm should be cheaper and more efficient, so pretty much better in every way.

 

[z0m3le]

That article sure did manage to ruffle a lot of feathers without saying anything at all.

It claims that they couldn't use Pascal due to timing issues, but does that make any sense to anyone?

Also, as far as 16nm Maxwell is concerned, 16nm processes have been available for a long time, right? At least most of this year? 16nmFF+ or 16nmFFC might not have been available as long, but in general, for a product being mass produced in late 2016/early 2017 what possible reason could there be to go 20nm? 16nm should be cheaper and more efficient, so pretty much better in every way.

This seems to be based on the final devkits being maxwell still, that doesn't mean final hardware won't be pascal or 16nm, especially because battery life is better than expected? Again it doesn't matter much but there isn't too much time until we know more about the retail units.

 

[Schnozberry]

That article sure did manage to ruffle a lot of feathers without saying anything at all.

It claims that they couldn't use Pascal due to timing issues, but does that make any sense to anyone?

Also, as far as 16nm Maxwell is concerned, 16nm processes have been available for a long time, right? At least most of this year? 16nmFF+ or 16nmFFC might not have been available as long, but in general, for a product being mass produced in late 2016/early 2017 what possible reason could there be to go 20nm? 16nm should be cheaper and more efficient, so pretty much better in every way.

The timing issue makes a lot of sense for me. If they had originally planned this for Holiday 2016, they may have chosen Maxwell because they would have been cutting it too close to their original planned release for Holiday 2016.

 

[BDGAME]

So, if that information is valid, I want to understand that part:

https://twitter.com/NWPlayer123/status/789116886109655041

Four ARM Cortex-A57 cores, max 2GHz
NVidia second-generation Maxwell architecture
256 CUDA cores, max 1 GHz, 1024 FLOPS/cycle
4GB RAM (25.6 GB/s, VRAM shared)
32 GB storage (Max transfer 400 MB/s)
USB 2.0 & 3.0
1280 x 720 6.2" IPS LCD
1080p at 60 fps or 4k at 30 fps max video output
Capcitance method, 10-point multi-touch

Is that the first machine to utilize that secondary generation or it's nothing new?

 

[Skittzo0413]

This seems to be based on the final devkits being maxwell still, that doesn't mean final hardware won't be pascal or 16nm, especially because battery life is better than expected? Again it doesn't matter much but there isn't too much time until we know more about the retail units.

The final devkits could be Maxwell on 16nm for all we know. The point is that article tells us nothing without any inside info about the process node.

The timing issue makes a lot of sense for me. If they had originally planned this for Holiday 2016, they may have chosen Maxwell because they would have been cutting it too close to their original planned release for Holiday 2016.

Well, firstly, if they were targeting Holiday 2016 then they clearly delayed it. Was the delay purely about software? If so, why not take advantage of the cheaper and more efficient manufacturing process anyway? Would they have built these chips months ago and had them sitting in a warehouse all this time?

And again, Maxwell doesn't mean 20nm. They could have chosen Maxwell on 16nm and it would be just as efficient as Pascal. Haven't 16nm processes been available for quite a while? Even possible for a Holiday 2016 launch?

So, if that information is valid, I want to understand that part:

Is that the first machine to utilize that secondary generation or it's nothing new?

I would assume that's referring to the 20nm process of the TX1 rather than the previous 28nm process.

 

[Donnie]

This seems to be based on the final devkits being maxwell still, that doesn't mean final hardware won't be pascal or 16nm, especially because battery life is better than expected? Again it doesn't matter much but there isn't too much time until we know more about the retail units.

Heavily customised Maxwell isn't it?, as in development of the chip began with Maxwell as the base, which obviously it would do considering development started more than two years ago. Development of Pascal also began with Maxwell as the base. Sounds more like Pascal and this chip are two different offs-hoots of Maxwell.

People talking about choosing to use Maxwell or Pascal need to remember this is a custom GPU, so they haven't chosen to use either. Obviously it uses the same shader units as Pascal, because so does Maxwell. The real question is what node have they used, that's something nobody has specified in any of these new "leaks".

 

[Lom1lo]

16nm Maxwell would be Pascal ? Thought Pascal is nothing more than a die shrink ?!

 

[z0m3le]

Heavily customised Maxwell isn't it?, as in development of the chip began with Maxwell as the base, development of Pascal also began with Maxwell as the base. Sounds more like Pascal and this chip are two offshoots of Maxwell.

Pascal and Maxwell are the same chips with some modifications, X1 for instance already got most of Pascal's features that I know of, and was a shrink to 20nm where as pascal is a modified maxwell at 16nm. IE Pascal is a made up architecture to sell GPUs to all those 970 and 980 customers again:

 

[Clessidor]

So, if that information is valid, I want to understand that part:



Is that the first machine to utilize that secondary generation or it's nothing new?

It seems like there exist two generations of Maxwell architecture. The second Generation is the more modern one.
So nothing new I think

 

[Donnie]

Pascal and Maxwell are the same chips with some modifications, X1 for instance already got most of Pascal's features that I know of, and was a shrink to 20nm where as pascal is a modified maxwell at 16nm. IE Pascal is a made up architecture to sell GPUs to all those 970 and 980 customers again:

Yeah that's my point. Pascal's development (which consists of a die shrink and slightly improved buffer compression) started with Maxwell as the base. Switch's GPU development also started with Maxwell as the base but is obviously different in its customisation, likely much more customised than Pascal.

So of course Switch's GPU isn't Pascal, and theirs no point in talking about timelines of Tegra availability and which GPU Nintendo chose between Maxwell and Pascal as they haven't chosen a off the shelf GPU They've had Nvidia develop a custom GPU which is something similar to Maxwell/Pascal but with its own customisations, on which node we don't know. We can guess logically 16nm.

 

[MuchoMalo]

Guys, it's 20nm. Battery life will be abysmal. Accept it.

Reminder here too, that article speaks about Maxwell AND 1TF power, while in here people have been discussing 750GF (at most!) and Pascal.

So, which ever is true, it's a win for us.

And please someone put that poor fellow out of his misery already!

EDIT. Yeah, my flips beats everyone else's flops... I'll just go lie down for a moment.

It's obviously talking about FP16, so effective performance is in-between 512 and 1024GFLOPS. It's the exact same range. Most likely, it'll be more like 384-435 FP32 undocked.

 

[z0m3le]

I know that's my point. Pascal's development (which consists of a die shrink and slightly improved buffer compression) started with Maxwell as the base. Switch's GPU also started with Maxwell as the base but is obviously different in its customisation, likely much more customised than Pascal was.

So of course Switch's GPU isn't Pascal, and there's no point in talking about timelines of Tegra availability and which GPU Nintendo chose between Maxwell and Pascal. They've had Nvidia develop a custom GPU which is something similar to Maxwell/Pascal but with its own customisations, on which node we don't know. We can guess logically 16nm.

This is where I am on this as well. People report on tech all the time without understanding it, or researching anything but the shallowest of details.

 

[Skittzo0413]

Guys, it's 20nm. Battery life will be abysmal. Accept it.

Why would that be the case if it's cheaper to go 16nm?

It's obviously talking about FP16, so effective performance is in-between 512 and 1024GFLOPS. It's the exact same range. Most likely, it'll be more like 384-435 FP32 undocked.

He then compares that 1TF number to the FP32 numbers of the PS4 and Scorpio, and then says that the Switch might be more powerful than the XB1, so who the hell knows what he's basing that on. Most of his conclusions are contradictory nonsense.

 

[z0m3le]

Guys, it's 20nm. Battery life will be abysmal. Accept it.



It's obviously talking about FP16, so effective performance is in-between 512 and 1024GFLOPS. It's the exact same range. Most likely, it'll be more like 384-435 FP32 undocked.

Pixel C is passively cooled X1 at 435GFLOPs FP32 and is a tablet with a similar area for heat displacement to Switch. I'd put that at the bare minimum tbh, it even uses 8 watts during the Manhattan benchmark for the entire system with a bigger, brighter, higher resolution screen and more CPU cores than Switch will have.

 

[Donnie]

Guys, it's 20nm. Battery life will be abysmal. Accept it..

According to who?, or are you joking here?

 

[Mokujin]

Also, as far as 16nm Maxwell is concerned, 16nm processes have been available for a long time, right? At least most of this year? 16nmFF+ or 16nmFFC might not have been available as long, but in general, for a product being mass produced in late 2016/early 2017 what possible reason could there be to go 20nm? 16nm should be cheaper and more efficient, so pretty much better in every way.

Some people are making such a big fuss about Maxwell vs. Pascal when in itself the differences are negligible architecture wise, fab node being 16nm or 20nm on the other side it's really a big deal.

Switch SoC being on 20nm would be the most negative news we could have because it's so much worse than 16nm, but after the confirmation that Switch can work at full clocks docked I'm quite comfortable knowing that 512 Glops it's an almost unmissable baseline in this mode.

The only way I see Nintendo / Nvidia keeping the design at 20nm is if the foundry were selling these suuuuuuper-cheap to clear stocks or amortize machinery, otherwise it makes little sense.

Reminder here too, that article speaks about Maxwell AND 1TF power, while in here people have been discussing 750GF (at most!) and Pascal.

So, which ever is true, it's a win for us.

And please someone put that poor fellow out of his misery already!

EDIT. Yeah, my flips beats everyone else's flops... I'll just go lie down for a moment.

This is a common mistake being done over and over, Tegra X1 is already a 1Tf chip but at fp16-half precission and while there is some optimization that can be done with these, the number with real meaning are fp32 flops (512).

That Venturebeat article looks like a mashup of the already known info with no extra real info added and some errors that are going to (already are) confuse people.

So in order of relevance I still think that the most relevants customizations vs Tegra X1 are:

1) 16 nm instead of 20nm
2) Memory setup
3) CPU setup

Maxwell of Pascal doesn't really matter that much.

 

[Markitron]

So if the specs in the OP are correct, what kind of power are we roughly looking at? Closer to an XB1 than a Wii-U?

Would you guys consider it more or less powerful that the Wii-U and Wii at their respective launches (when compared with the competition).

 

[Interfectum]

So if the specs in the OP are correct, what kind of power are we roughly looking at? Closer to an XB1 than a Wii-U?

Would you guys consider it more or less powerful that the Wii-U and Wii at their respective launches (when compared with the competition).

Some insiders have suggested the gap between Switch and Xbone is less than the gap between Xbone and PS4. Take with a grain of salt though.

 

[FyreWulff]

So if the specs in the OP are correct, what kind of power are we roughly looking at? Closer to an XB1 than a Wii-U?

Would you guys consider it more or less powerful that the Wii-U and Wii at their respective launches (when compared with the competition).

It's going to be closer to an XB1 either way. There's a huge gap between what the U had it in it vs what's available in mobile form factors.