Support NeoGAF

[Skittzo0413]

Power efficiency and how high the clocks an architecture can reach are two very different things. Otherwise, we would have seen much higher clocks out of AMD's Polaris cards than we actually did.

The way I understand clock speed of processing components is that the upper limit has to do with thermal constraints, as the component getting too hot would eventually lead to electrical failure. From the fact that a 16nm Maxwell CUDA core can generate the same amount of heat as a 20nm core at a higher clock rate due to its increased power efficiency, it would make sense that it could be clocked higher specifically because of that increased efficiency.

If that's not the case, then what causes the upper limit of a clock rate?

 

[blu]

Accepted terms (so we can avoid arguments over t's and i's):

flop - floating point operation
FLOPS - Floating Point Operations Per Second
fp16 - 16-bit floating-point format (AKA half precision) - 11-bit mantissa, 5-bit exponent, 1 sign bit.
fp32 - 32-bit floating-point format (AKA single precision) - 24-bit mantissa, 8-bit exponent, 1 sign bit
fp64 - 64-bit floating-point format (AKA double precision) - 54-bit mantissa, 10-bit exponent, 1 sign bit

Did I miss something that could potentially become a source of misunderstanding?

 

[MuchoMalo]

I would be dissapointed if Nintendo went 20nm process, not much for horse power implicancies but seeing that will be stuck with an old process that likely will increase the production cost just like Wii U.

I think that you guys need to stop focusing on these numbers and just wourry about how the battery life ends up. I don't think that Nintendo would have gone for more performance over more battery life on 16nmFF. (assuming it's on 20nm) If the battery life is still decent, then we should just take what we get.

 

[Doctre81]

The article says nothing about FP16 or FP32, and they compare that number of 1TF to the 6TF of Scorpio. The point is, the author of that article (assuming he saw the number for FP16) clearly doesn't have enough technical knowledge to draw conclusions based on microarchitecture.

The article doesn't mention FP16 because the author doesn't even know what that is.

 

[Clessidor]

How can we give legitimacy to the Venture article when it says that the switch will perform at 1 teraflop.

The thing is NathanDrake and Emily Rogers also said they heared that the final Devkits have Maxwell. The person who wrote the article might have a lack technical understanding and it's horribly written.
But at least the rumour that the Switch will be Maxwell-based seems to be likely. It's also the only thing he claims to got from soures outside. Stuff like it will perfom at 1 teraflop is something he pulled out of nowhere.

So it's fair to assume that the one information, which is true is that Switch might have a Maxwell-based GPU or at least to talk about it.

EDIT: fixed it. thx for pointing it out

 

[Skittzo0413]

Accepted terms (so we can avoid arguments over t's and i's):

flop - floating point operation
FLOPS - Floating Point Operations Per Second
fp16 - 16-bit floating-point format (AKA half precision) - 11-bit mantissa, 5-bit exponent, 1 sign bit.
fp32 - 32-bit floating-point format (AKA single precision) - 24-bit mantissa, 8-bit exponent, 1 sign bit
fp64 - 64-bit floating-point format (AKA double precision) - 54-bit mantissa, 10-bit exponent, 1 sign bit

Did I miss something that could potentially become a source of misunderstanding?

The OP says "1024 FLOPS/cycle" - FLOPS/cycle is a nonsense unit, right?

The article doesn't mention FP16 because the author doesn't even know what that is.

That's exactly my point, yeah. If the author is saying 1TF and comparing that to Scorpio's 6TF he clearly doesn't understand this stuff, so why are we taking any of it seriously?

The thing is NathanDrake and Emily Rogers also said they heared that it will be Maxwell. The person who wrote the article might have a lack technical understanding and it's horribly written.
But at least the rumour that the Switch will be Maxwell-based seems to be likely. It's also the only thing he claims to got from soures outside. Stuff like it will perfom at 1 teraflop is something he pulled out of nowhere.

So it's fair to assume that the one information, which is true is that Switch will have a Maxwell-based GPU or at least to talk about it.

Did Nate say that? He said the final devkits are Maxwell based, but he still heard the final unit will be Pascal based as of the summer, right? Unless I missed something. I believe Emily also was vouching for the devkit's architecture, not the final unit's.

 

[KojiKnight]

I would be dissapointed if Nintendo went 20nm process, not much for horse power implicancies but seeing that will be stuck with an old process that likely will increase the production cost just like Wii U.

Nothing says they'd be "stuck" on it. Consoles upgrade to smaller processes all the time. It's what usually makes "slim" consoles possible.

 

[Doctre81]

The thing is NathanDrake and Emily Rogers also said they heared that it will be Maxwell. The person who wrote the article might have a lack technical understanding and it's horribly written.
But at least the rumour that the Switch will be Maxwell-based seems to be likely. It's also the only thing he claims to got from soures outside. Stuff like it will perfom at 1 teraflop is something he pulled out of nowhere.

So it's fair to assume that the one information, which is true is that Switch will have a Maxwell-based GPU or at least to talk about it.

They said in the dev kits. Emily said she heard final consumer hardware will be "similar" pascal is based of mxwell so that makes it similar. Nate heard pascal in final.

 

[th3sicknness]

The thing is NathanDrake and Emily Rogers also said they heared that it will be Maxwell. The person who wrote the article might have a lack technical understanding and it's horribly written.
But at least the rumour that the Switch will be Maxwell-based seems to be likely. It's also the only thing he claims to got from soures outside. Stuff like it will perfom at 1 teraflop is something he pulled out of nowhere.

So it's fair to assume that the one information, which is true is that Switch will have a Maxwell-based GPU or at least to talk about it.

Aren't their reports based off dev kits and not retail units? Nate confirmed that the most recent final kit was Maxwell he didn't mention anything about retail units being the same ?

 

[LordKano]

Is it possible to change the architecture in such a short amount of time ? Would it change something if, say, it was Maxwell 16nm in devkits, but Pascal 16nm in retail ?

 

[Hermii]

Aren't their reports based off dev kits and not retail units? Nate confirmed that the most recent final kit was Maxwell he didn't mention anything about retail units being the same ?

Most recent final kit? Doesn't final mean final as in there will be no more kits?

 

[AlStrong]

The OP says "1024 FLOPS/cycle" - FLOPS/cycle is a nonsense unit, right?

Yes.

 

[Doctre81]

Is it possible to change the architecture in such a short amount of time ? Would it change something if, say, it was Maxwell 16nm in devkits, but Pascal 16nm in retail ?

the only Big difference really between a 16nm maxwell and 16nm pascal tegra would be 64bit bus vs 128 bit. Unless I'm missing something.

 

[Skittzo0413]

Is it possible to change the architecture in such a short amount of time ? Would it change something if, say, it was Maxwell 16nm in devkits, but Pascal 16nm in retail ?

The way I understand it, a custom Maxwell chip at 16nm would be pretty much the same as a Pascal chip, with a few minor tweaks that don't really affect a gaming device. So what I think is happening is that the final devkits ARE Maxwell based custom chips at 16nm, and the Venture Beat article is essentially worthless. This would explain why Nate has heard Pascal, because a lot of people -even tech minded people here- think Maxwell only exists on 20nm (and 28nm) and Pascal is the only way to get Tegra on a 16nm chip.

the only Big difference really between a 16nm maxwell and 16nm pascal tegra would be 64bit bus vs 128 bit. Unless I'm missing something.

I would imagine the RAM bus is something which would be part of the customizations being made.

 

[Hermii]

the only Big difference really between a 16nm maxwell and 16nm pascal tegra would be 64bit bus vs 128 bit

Difference between 16nm and 16nm pascal or maxwell would be trivial. Maxwell implies it's 20nm.

 

[blu]

The OP says "1024 FLOPS/cycle" - FLOPS/cycle is a nonsense unit, right?

It is nonsense, but the OP either did not know the difference between flops (lower case) and FLOPS (upper case), or just did not care. If we assume it meant flops/clock it all falls into place.

 

[Doctre81]

I think that the dev kits say "maxwell" because pascal was not known. In otherwords I think it's pascal technically they just didn't call it that at the time.

 

[AlStrong]

It's is nonsense

No u.

edit: curse your edit.

 

[Skittzo0413]

It is nonsense, but the OP either did not know the difference between flops (lower case) and FLOPS (upper case), or just did not care. If we assume it meant flops/clock it all falls into place.

So 1024flops/cycle (or /clock) would equate to 1TFLOPS, likely FP16?

No u.

edit: curse your edit.

Ha I actually thought it was intentional ("it's is" is the same type of meaning as "floating point operations per second per clock"). Just a happy accident then.

 

[AlStrong]

So 1024flops/cycle (or /clock) would equate to 1TFLOPS, likely FP16?

As I mentioned above, 1024 flop/clk is what you get from 256 CUDA cores with double rate fp16.

 

[Skittzo0413]

As I mentioned above, 1024 flop/clk is what you get from 256 CUDA cores with double rate fp16.

Gotcha, thank you. I was uncertain of the differences between flops and FLOPS when you posted that.

Spoiler

Why must we make everything into an acronym...

 

[Doctre81]

Here is a scary thought.

What if the reason why they are using 20nm is because they aren't manufacturing the chips in 2017...


They already manufactured 40 million of them way back in 2015 and are just waiting for the games to be finished.

Just kidding.

I hope.

 

[Schnozberry]

Here is a scary thought.

What if the reason why they are using 20nm is because they aren't manufacturing the chips in 2017...


They already manufactured 40 million of them way back in 2015 and are just waiting for the games to be finished.

Just kidding.

I hope.

Zero percent chance of this. Based on information we already know (USB 3.1 support for video over USB C), it's not a standard Tegra X1 that is just laying around.

 

[Doctre81]

Zero percent chance of this. Based on information we already know (USB 3.1 support for video over USB C), it's not a standard Tegra X1 that is just laying around.

I know. Just wanted to spill some coffee lol. Nintendo already confirmed the systems innards were still being developed earlier this year.

 

[Vena]

Zero percent chance of this. Based on information we already know (USB 3.1 support for video over USB C), it's not a standard Tegra X1 that is just laying around.

I mean, that basic point would also torpedo any idea of this being some left-over 20nm X1, but people keep floating that idea around.

 

[blu]

No u.

edit: curse your edit.

Typing gibberish is perfectly legit at my age /sips tea

 

[Schnozberry]

I mean, that basic point would also torpedo any idea of this being some left-over 20nm X1, but people keep floating that idea around.

Well, it doesn't rule out a custom chip fabbed on 20nm, but it does rule out a bog standard Jetson TX1 like this OP details.

 

[LordKano]

I mean, that basic point would also torpedo any idea of this being some left-over 20nm X1, but people keep floating that idea around.

Couldn't it be custom chip based on X1 but still 20nm ? That seems to be what most people think here.

 

[Vena]

Well, it doesn't rule out a custom chip fabbed on 20nm, but it does rule out a bog standard Jetson TX1 like this OP details.

I meant the "stockpile of 20nm X1s" that people like to quote.

Couldn't it be custom chip based on X1 but still 20nm ? That seems to be what most people think here.

Yes, but that wouldn't be from a "stockpile", that'd be a new chip being fabbed... new.

 

[blu]

Well, it doesn't rule out a custom chip fabbed on 20nm, but it does rule out a bog standard Jetson TX1 like this OP details.

I don't know why people are even remotely entertaining the idea of a stock TX1 in the Switch. It's like that other thread I shall not name where an amount of stupidity I did not suspect existed on gaf poured out. In waterfalls.

 

[kIdMuScLe]

I don't know why people are even remotely entertaining the idea of a stock TX1 in the Switch. It's like that other thread I shall not name where an amount of stupidity I did not suspect existed on gaf poured out. In waterfalls.

Lol I read some posts over on 4chan that they fed that fake info to the site just to see the reactions

 

[Schnozberry]

I don't know why people are even remotely entertaining the idea of a stock TX1 in the Switch. It's like that other thread I shall not name where an amount of stupidity I did not suspect existed on gaf poured out. In waterfalls.

Whenever a negative Nintendo thread comes around, the amount of mouth breathers that show up to dogpile on one thing or another is astounding.

 

[Easy_D]

I don't know why people are even remotely entertaining the idea of a stock TX1 in the Switch. It's like that other thread I shall not name where an amount of stupidity I did not suspect existed on gaf poured out. In waterfalls.

This is the forum that creamed its pants repeatedly over GDDR5, I'd wager a majority did not know what it meant other than 5 being a higher number than 3.

 

[Schnozberry]

This is the forum that creamed its pants repeatedly over GDDR5

Seriously, it's like 2 DDR's better than GDDR3.

 

[magash]

Whenever a negative Nintendo thread comes around, the amount of mouth breathers that show up to dogpile on one thing or another is astounding.

In this case Nvidia stated without mincing words that the switches solution is based on a customized Tegra solution. I think Nvidia even stated that they invested a couple of hundreds of hours in man years developing switches solution. I believe because its Nintendo some people have a preconceived bias.

 

[Doctre81]

Couldn't it be custom chip based on X1 but still 20nm ? That seems to be what most people think here.

A custom 20nm would cost more money to produce than a custom 16nm one.

 

[Hermii]

A custom 20nm would cost more money to produce than a custom 16nm one.

We have no idea what kind of deal Nintendo and Nvidia negotiated.

 

[th3sicknness]

A custom 20nm would cost more money to produce than a custom 16nm one.

How do you know that with certainty?

 

[Doctre81]

How do you know that with certainty?

Because of the smaller parts. Console makers don't just come out slimmer designs later down the line because they think they look better. They make smaller designs because eventually it becomes cheaper to do that over the originals.

 

[Schnozberry]

A custom 20nm would cost more money to produce than a custom 16nm one.

For whom? Nvidia had existing wafer contracts on 20nm, and rather than pay looming penalties for backing out of their commitments, they may have offered them to Nintendo at a very reasonable price. AMD had to pay $33M in penalties to cancel their 20nm chips, and had to spend $320M to amend their wafer agreements with GlobalFoundries.

It's a win-win for Nintendo, because if the Switch is a success the die shrink after the 20nm commitments are exhausted will allow them to release a revision that runs cooler with significantly increased battery life.

 

[Doctre81]

For whom? Nvidia had existing wafer contracts on 20nm, and rather than pay looming penalties for backing out of their commitments, they may have offered them to Nintendo at a very reasonable price. AMD had to pay $33M in penalties to cancel their 20nm chips, and had to spend $320M to amend their wafer agreements with GlobalFoundries.

It's a win-win for Nintendo, because if the Switch is a success the die shrink after the 20nm commitments are exhausted will allow them to release a revision that runs cooler with significantly increased battery life.

Nintendo may have offered to help with that bill if they thought it would benefit them financially in the long run. I doubt switch is 20nm.

 

[EDarkness]

I don't know why people are even remotely entertaining the idea of a stock TX1 in the Switch. It's like that other thread I shall not name where an amount of stupidity I did not suspect existed on gaf poured out. In waterfalls.

Seriously. What the hell is going on in here? I thought people were being reasonable before.

 

[Schnozberry]

Nintendo may have offered to help with that bill if they it would benefit them financially in the long run. I doubt switch is 20nm.

How does that benefit Nintendo financially in the long run? It would be more advantageous to buy up the 20nm wafers on the cheap and make use of them, rather than offer to pay to bail Nvidia out of a bad deal.

 

[Doctre81]

How does that benefit Nintendo financially in the long run? It would be more advantageous to buy up the 20nm wafers on the cheap and make use of them, rather than offer to pay to bail Nvidia out of a bad deal.

We are also assuming that nvidia had a bunch of 20nm wafers just sitting there. That may not even be true. Nvidia was not please with 20nm at least initially so maybe they didn't buy as many as thought.

 

[Screamapillar]

This looks/ seems about right? Modest specs, nothing seems too poor, or too unrealistically good. This thing's gotta be able to run graphics superior to Wii U on battery power at 720p in portable mode.

 

[Doctre81]

This looks/ seems about right? Modest specs, nothing seems too poor, or too unrealistically good. This thing's gotta be able to run graphics superior to Wii U on battery power at 720p in portable mode.

what's off about it is that 16nm would have offered far superior battery life and power consumption.

 

[Hermii]

Who has actually said it's a stock tx1? What I seen has been entertained is a semicustom design based on tx1 or at least based on 20nm
Maxwell.

 

[th3sicknness]

For whom? Nvidia had existing wafer contracts on 20nm, and rather than pay looming penalties for backing out of their commitments, they may have offered them to Nintendo at a very reasonable price. AMD had to pay $33M in penalties to cancel their 20nm chips, and had to spend $320M to amend their wafer agreements with GlobalFoundries.

It's a win-win for Nintendo, because if the Switch is a success the die shrink after the 20nm commitments are exhausted will allow them to release a revision that runs cooler with significantly increased battery life.

Didnt the same already occur for Nvidia? Nvidia got out of 20nm business because it was shit just as AMD did .. why would they sell it to Nintendo and let them tinker with it and put out a product that would reflect badly on them?

 

[AzaK]

I'm expecting around 450 GFLOPS while it's in handheld mode and 650-700 GFLOPS docked; which in my opinion is excellent for a portable. The Wii U is 350 GFLOPS and I've always been satisfied with how Nintendo games look because of their great art direction, the modern architecture should also help separate the lines a little more. Anyone who is looking at the Switch to play anything more than first party and indie games like Minecraft and Terraria is probably setting themselves up for disappointment though.

Wii U is 176

 

[EDarkness]

Who has actually said it's a stock tx1? What I seen has been entertained is a semicustom design based on tx1 or at least based on 20nm
Maxwell.

Between this thread and the other thread, it's come up a lot. I think speculation without anything to go on is just sending this line of thinking around in circles. Now Nvidia had some 20nm wafers just laying around....