Support NeoGAF

[Graphics Horse]

They say about the "enhancer" that "It connects to the back of the main unit motherboard via some sort of PCI bridge.", which means it's a separate device which the Switch attaches to. The logic would be that it replaces the dock, hence why the Switch attaches to it, and why it has video out, no battery and a built-in power supply. That isn't the form factor of the dev kit they're describing, but early dev kits rarely resemble the final form factor of a piece of gaming hardware.

Which still describes the Dev unit photographed to a tee. So we are relying on the memory and processor on the second board being a prototype future dock, rather than just to aid development. I guess it's possible as a 4k version as the guy speculates. Would the USB be good enough for the CPU to access the other Ram at native speeds?

 

[Skittzo0413]

I'm completely ignorant but I'm going to speculate anyway.

What if the everything in the Foxconn leak is true, but it's still on 20nm? They are running the A57s at 1.78 Ghz in a stress test, and at the same time they are building a SCD with a 1060-ish GPU. Could it not be that they are testing the possibility of a dock that comes with extra GPU hardware and upclocks the CPU significantly in order to run PS4/Xbone-caliber software?

So the SCD has GPU only, the Switch has Eurogamer clocks in its release state, but if you buy the SCD dock it will utilize the SCD and upclock its own CPU to levels that cannot be sustained in portable mode.

I think the issue with this is that the Shield TV can't even maintain the clocks mentioned by the leaker for more than a few minutes. It automatically throttles those clocks down to around the levels cited by DF, and has far more volume/surface area than the Switch, and therefore a far better cooling arrangement.

So if the Shield TV can't even manage those clocks for an hour with a 20nm SoC, then how could the Switch do so for 8 days?

 

[Rodin]

Well, the patent did specifically refer to the fan running (at a lower RPM, afaik) while in handheld mode, so it shouldn't be that surprising. Thinking about it a little bit more, though, it's more likely that the fan isn't specifically tied to whether the unit is docked or not, but rather the temperature of the SoC (like virtually all other cooling solutions). The fan may remain on for a period of time after being undocked, until the SoC temp drops down below a particular threshold. Alternatively, an intensive game running in portable mode may cause the fan to kick in occasionally for short bursts if the temperature picks up. Similarly, the fan might actually stay off in docked mode for some games, such as 1, 2, Switch, which is unlikely to stress the SoC in either mode.

I suggested many times that the fan could act like in modern desktop GPUs. My MSI GTX 970 has a 0 rpm mode in idle and with less intensive games, but it kicks in when demanding titles push it more and the temperature goes up. I imagine that the little fan in the Switch could do the exact same thing to avoid throttling.

I'm personally taking the 16nm claim as the leaker's assumption, as I don't see any way he would have had that info. Die sizes could obviously be measured, and the fact that they're made in Taiwan would be printed on the chip, but there's no reason for the fabrication process to be printed anywhere (especially if the manufacturer isn't even detailed).

Agreed, i think he's assuming this based on the clocks he read on the screen, which would likely be too high for a 20-28nm SoC.

Reading through Time's interview with Takahashi and Koizumi, I couldn't help but notice this particular comment by Koizumi:

Granted, he preceded this by commenting about the removable controllers, so he's likely talking about add-ons like the joycon variant with a proper d-pad we saw in the patent, but it's still a very interesting comment in the context of this thread. It also reiterates Nintendo's new approach of holding off on product reveals until as close to launch as possible. If Nintendo does release some kind of upgraded GPU dock, then I wouldn't expect to hear about it until perhaps a month or two before it hits shelves (although like PS4Pro, it would almost certainly leak beforehand).

They say about the "enhancer" that "It connects to the back of the main unit motherboard via some sort of PCI bridge.", which means it's a separate device which the Switch attaches to. The logic would be that it replaces the dock, hence why the Switch attaches to it, and why it has video out, no battery and a built-in power supply. That isn't the form factor of the dev kit they're describing, but early dev kits rarely resemble the final form factor of a piece of gaming hardware.

This is pretty exciting, i'm really curious to see if this "enhancer" device will actually be released on the market. I think it would be smart, it's basically an optional accessory that replaces your dock to boost power for people who want better docked performances (for example people with a 4K TV), and it shouldn't be too hard for Nintendo to make 4K patches for existing major titles (like Zelda, Mario, MK, Xenoblade 2 or Splatoon) and support 2-3 different res going forward. Seems like the machine was designed for this, the SCD patent has been around for a while, not to mention that they unified their development teams. It's still much easier than making Smash 3DS and Wii U, Super Mario Maker 3DS and Wii U, Yoshi 3DS and Wii U, Mario kart 7 and 8, etc...

 

[z0m3le]

I think the issue with this is that the Shield TV can't even maintain the clocks mentioned by the leaker for more than a few minutes. It automatically throttles those clocks down to around the levels cited by DF, and has far more volume/surface area than the Switch, and therefore a far better cooling arrangement.

So if the Shield TV can't even manage those clocks for an hour with a 20nm SoC, then how could the Switch do so for 8 days?

And this is why the 20nm switch doesn't make sense, and why I'd give the leaker the benefit of the doubt with 16nm, not to mention that a 4k dock being joined to a 1ghz A57 quad core is way too lopsided to be a thing for a company like Nintendo, especially after they released the new 3ds with the much faster cpu than the original.

 

[Skittzo0413]

And this is why the 20nm switch doesn't make sense, and why I'd give the leaker the benefit of the doubt with 16nm, not to mention that a 4k dock being joined to a 1ghz A57 quad core is way too lopsided to be a thing for a company like Nintendo, especially after they released the new 3ds with the much faster cpu than the original.

Yeah while I was typing that I up I sorta convinced myself to move closer to 60/40 on 16nm being the final configuration. If the Shield TV does indeed throttle to the clocks reported in the DF article like MDave discovered, then I don't see how the Switch could possibly be tested at those clocks for 8 days without it being 16nm.

The only alternate explanation there that I could think of would be that the leaker fabricated that part of his story. Which doesn't seem likely at this point.

 

[Piscus]

I think the issue with this is that the Shield TV can't even maintain the clocks mentioned by the leaker for more than a few minutes. It automatically throttles those clocks down to around the levels cited by DF, and has far more volume/surface area than the Switch, and therefore a far better cooling arrangement.

So if the Shield TV can't even manage those clocks for an hour with a 20nm SoC, then how could the Switch do so for 8 days?

Could Nintendo /Nvidia potentially modify the chip so it won't throttle? I'm not sure how that works... but could they separate CPU and GPU from each other so that they can be clocked independently of what the other is doing heat wise and just use the fan to cool?

 

[Skittzo0413]

Could Nintendo /Nvidia potentially modify the chip so it won't throttle? I'm not sure how that works... but could they separate CPU and GPU from each other so that they can be clocked independently of what the other is doing heat wise and just use the fan to cool?

From what I understand, this isn't possible.

The Tegra chip is a SoC, which means the CPU and GPU are both located on the same die. The distance between the CPU and GPU is on the order of milimeters, and the heat generated will spread much further than that. Processors in most mobile devices will throttle to protect the battery, but like in the case of the Shield TV, also in order to protect the actual circuitry and components from frying/melting under the excess heat.

So when these chips are being developed, they are tested to see the maximum possible heat loads they can take without being damaged.

Since the TX1 in the Shield TV does indeed throttle to levels similar to those from the DF article, it would stand to reason that running the clocks at 2GHz for the CPU and 1GHz for the GPU for too long would cause electrical failure and destroy the device.

Yet here we have a report of the Switch being run at 1.78GHz for the CPU and 921MHz for the GPU for a full 8 days, when the TX1 in the Shield can only run for a matter of minutes before either the CPU needs to be throttled to 1GHz or the GPU needs to be throttled to 768MHz, or some combination in the middle, in order to protect the components from failure.

And then on top of that, the Switch is a much, much thinner and smaller device and therefore has a less effective cooling arrangement than the Shield TV, since cooling largely depends on surface area.

It seems to me that (again, if the Shield TV does indeed throttle as MDave discovered) a single SoC in the Switch could not be run at those clocks for 8 days if it was a TX1 on a 20nm process.

 

[forknknife]

Don't know were to post this so I'll do it here.

The switch can apparently project in both the TV and the switch screen at the same time via USB-C. That's interesting right?

 

[calc84maniac]

And this is why the 20nm switch doesn't make sense, and why I'd give the leaker the benefit of the doubt with 16nm, not to mention that a 4k dock being joined to a 1ghz A57 quad core is way too lopsided to be a thing for a company like Nintendo, especially after they released the new 3ds with the much faster cpu than the original.

Even on 20nm, shouldn't they be able to run the CPU at 2GHz while using only the external GPU? Though there is still the stress test to consider...

Don't know were to post this so I'll do it here.

https://i.redditmedia.com/9ELk4R2Ng6cBg3Yl3xLrs1m3_eWAqCTwzUTQCVFjQ6s.jpg?w=1023&

The switch can apparently project in both the TV and the switch screen at the same time via USB-C. That's interesting right?

Broken link?

 

[nightside]

From what I understand, this isn't possible.

The Tegra chip is a SoC, which means the CPU and GPU are both located on the same die. The distance between the CPU and GPU is on the order of milimeters, and the heat generated will spread much further than that. Processors in most mobile devices will throttle to protect the battery, but like in the case of the Shield TV, also in order to protect the actual circuitry and components from frying/melting under the excess heat.

So when these chips are being developed, they are tested to see the maximum possible heat loads they can take without being damaged.

Since the TX1 in the Shield TV does indeed throttle to levels similar to those from the DF article, it would stand to reason that running the clocks at 2GHz for the CPU and 1GHz for the GPU for too long would cause electrical failure and destroy the device.

Yet here we have a report of the Switch being run at 1.78GHz for the CPU and 921MHz for the GPU for a full 8 days, when the TX1 in the Shield can only run for a matter of minutes before either the CPU needs to be throttled to 1GHz or the GPU needs to be throttled to 768MHz, or some combination in the middle, in order to protect the components from failure.

And then on top of that, the Switch is a much, much thinner and smaller device and therefore has a less effective cooling arrangement than the Shield TV, since cooling largely depends on surface area.

It seems to me that (again, if the Shield TV does indeed throttle as MDave discovered) a single SoC in the Switch could not be run at those clocks for 8 days if it was a TX1 on a 20nm process.

But I guess if those where the clocks shouldn't we already noticed in the games shown? Unless Nintendo is purposedely launching an underclocked machine? I don't know.

 

[Oresama]

*Tech speak too intimidating*

 

[runnin_blue]

Don't know were to post this so I'll do it here.

https://i.redditmedia.com/9ELk4R2Ng6cBg3Yl3xLrs1m3_eWAqCTwzUTQCVFjQ6s.jpg?w=1023&

The switch can apparently project in both the TV and the switch screen at the same time via USB-C. That's interesting right?

Your link doesn't work (on mobile anyway).

 

[blu]

A 4k dock just doesn't sound like Nintendo, but I can't imagine what else it could be...

A bespoke VR SCD?

 

[Skittzo0413]

But I guess if those where the clocks shouldn't we already noticed in the games shown? Unless Nintendo is purposedely launching an underclocked machine? I don't know.

No, why would we notice? It's a 20% increase in GPU power when docked, and we don't know if there's a comparable boost to the undocked clocks, even though it would seem likely. That's not something that would be noticeable at all.

~100% more CPU power also wouldn't be noticeable from a graphics standpoint, though it could be beneficial to framerate in some games.

Your link doesn't work (on mobile anyway).

Yeah it doesn't work on desktop either.

A bespoke VR SCD?

That could be, I guess. It still doesn't solve the problem of the 720p screen but if paired with a new screen in the HMD then I guess it could work.

I still think there is something in the works with the wireless local cloud aspect of the SCD patent. That seems like a very different, Nintendo way to do something about increased processing. But who knows.

 

[forknknife]

God damn it i can't make the image show

Here is the link

https://imgur.com/gallery/94wOk

 

[calc84maniac]

God damn it i can't make the image show

Here is the link

https://imgur.com/gallery/94wOk

Yeah, that needs more context. How do we know that TV isn't connected to a different Switch?

 

[Mokujin]

God damn it i can't make the image show

Here is the link

https://imgur.com/gallery/94wOk

It's not the same image but the same menu.

 

[ShadowFox08]

We need all the extra gpu we can get. 20% more gpu could help with the resolution(notably if its in the borderline), texture, polygons, or frame rate stability(those 3-6 more frames from Xbox one s to original Xbox comes to mind.

A good CPU not only is great for frame rate stability and speed of game, but physics, sound processing and A.I., and more objects on screen as well. 100% more CPU power would definitely be noticeable.

 

[Lexxism]

Is this the equation to get the gflops? 2 x cores x mhz / 1000 = gflops?

Based on EuroGamer. Docked: 2 x 256 x 768 /1000 = 393 gflops?

How about on foxconn? What's the rumoured mhz? Is it the same core?

 

[Buggy Loop]

nevermind

 

[Skittzo0413]

We need all the extra gpu we can get. 20% more gpu could help with the resolution(notably if its in the borderline), texture, polygons, or frame rate stability(those 3-6 more frames from Xbox one s to original Xbox comes to mind.

A good CPU not only is great for frame rate stability and speed of game, but physics, sound processing and A.I., and more objects on screen as well. 100% more CPU power would definitely be noticeable.

But noticeable compared to what? How would we be able to tell from the games shown at the various events whether it uses the DF CPU clocks or the Foxconn CPU clocks? We have no baseline for comparison here.

Is this the equation to get the gflops? 2 x cores x mhz / 1000 = gflops?

Based on EuroGamer. Docked: 2 x 256 x 768 /1000 = 393 gflops?

How about on foxconn? What's the rumoured mhz? Is it the same core?

921MHz. No one knows or claims to know the core configuration, though it's still likely to be 2SMs so 256 cores. So it would be 471 GFlops.

 

[lwilliams3]

But noticeable compared to what? How would we be able to tell from the games shown at the various events whether it uses the DF CPU clocks or the Foxconn CPU clocks? We have no baseline for comparison here.

Yeah. Either way, as Matt put it, the Switch is not as "CPU bottlenecked" as the other consoles. It will interesting to see how this will affect the games that takes full advantage of the hardware.

 

[ShadowFox08]

But noticeable compared to what? How would we be able to tell from the games shown at the various events whether it uses the DF CPU clocks or the Foxconn CPU clocks? We have no baseline for comparison here.



921MHz. No one knows or claims to know the core configuration, though it's still likely to be 2SMs so 256 cores. So it would be 471 GFlops.

hypothetically speaking

 

[z0m3le]

Yeah. Either way, as Matt put it, the Switch is not as "CPU bottlenecked" as the other consoles. It will interesting to see how this will affect the games that takes full advantage of the hardware.

Well Eurogamer's clocks would make Matt wrong, the A57 quad core even if all 4 cores are for games at 1ghz is only as capable as 4 PS4 cores, not 6 and a half like developers have with PS4.

 

[BuggyMike]

Well Eurogamer's clocks would make Matt wrong, the A57 quad core even if all 4 cores are for games at 1ghz is only as capable as 4 PS4 cores, not 6 and a half like developers have with PS4.

How would 4 A72 cores clocked at 1.78GHZ compare to PS4/X1?

 

[z0m3le]

How would 4 A72 cores clocked at 1.78GHZ compare to PS4/X1?

They would be faster than the ps4 overall and roughly on par with the ps4 pro's 8 cores, also they would be twice as fast in single threaded performance.

It's worth noting that development is often cpu bottlenecked and switch has been said to be the easiest of the 3 to develop on, certainly tools would offer an advantage but they will only get you so far and there is much more power, ram and bandwidth on other systems, so the comment is curious, especially given eurogamer's clocks.

 

[Hermii]

Yeah while I was typing that I up I sorta convinced myself to move closer to 60/40 on 16nm being the final configuration. If the Shield TV does indeed throttle to the clocks reported in the DF article like MDave discovered, then I don't see how the Switch could possibly be tested at those clocks for 8 days without it being 16nm.

The only alternate explanation there that I could think of would be that the leaker fabricated that part of his story. Which doesn't seem likely at this point.

If I remember oorrectly the shield does not throttle the CPU and it's clicked far higher than the switch.

 

[z0m3le]

If I remember oorrectly the shield does not throttle the CPU and it's clicked far higher than the switch.

It is not clocked much higher than the switch in this Foxconn leak for 8 days, also the gpu doesn't throttle according to the Foxconn leak even though it is very close to the x1's 1ghz, his point was given the lack of cooling compared to the X1, that what we know of the chip, it should have throttled at those clocks on 20nm. (the clocks being 1.78ghz and 921mhz which this device ran at without throttling for 8 days)

 

[lwilliams3]

Well Eurogamer's clocks would make Matt wrong, the A57 quad core even if all 4 cores are for games at 1ghz is only as capable as 4 PS4 cores, not 6 and a half like developers have with PS4.

In comparison to its GPU, Switch's CPU would be still a bit higher in performance. (50-75% of the PS4s CPU vs less than 40% of the GPU power.)

 

[Skittzo0413]

If I remember oorrectly the shield does not throttle the CPU and it's clicked far higher than the switch.

According to MDave's experiment here a couple months back the Shield TV does indeed throttle either the CPU or GPU when the other is at max load (unless that was debunked or otherwise proven wrong, in which case ignore all of this). And the 2GHz CPU in the Shield TV is really not that much higher than the Foxconn clock of 1.78GHz.

Basically-

Shield TV: 2GHz CPU, 1GHz GPU, reportedly throttles (to as low as 768MHz for GPU I recall) after a few minutes at max clocks

Switch Foxconn: 1.78GHz CPU, 921MHz GPU, reportedly ran at these clocks for 8 days straight

IF all of the above is true (also taking into account Switch's far smaller surface area for cooling) then the only conclusion I believe that is possible is that the Switch's SoC uses far less power, which would only be reasonably possible on 16nm.

 

[Hermii]

According to MDave's experiment here a couple months back the Shield TV does indeed throttle either the CPU or GPU when the other is at max load (unless that was debunked or otherwise proven wrong, in which case ignore all of this). And the 2GHz CPU in the Shield TV is really not that much higher than the Foxconn clock of 1.78GHz.

Basically-

Shield TV: 2GHz CPU, 1GHz GPU, reportedly throttles (to as low as 768MHz for GPU I recall) after a few minutes at max clocks

Switch Foxconn: 1.78GHz CPU, 921MHz GPU, reportedly ran at these clocks for 8 days straight

IF all of the above is true (also taking into account Switch's far smaller surface area for cooling) then the only conclusion I believe that is possible is that the Switch's SoC uses far less power, which would only be reasonably possible on 16nm.

Foxconn clocks at 16nm sounds great and plausible, but Im not quite ready believe Eurogamer got it wrong. 20nm at eurogamer clocks sounds just as plausible.

What will it take for us to know for sure? Another leak? Can we tell from a die shot?

 

[Rodin]

Well Eurogamer's clocks would make Matt wrong

I don't think so, Matt simply said that Switch wouldn't be as CPU limited as PS4, which, considering it's 157gflops undocked and 393-472 when docked, it's only obvious, even with 4 A57 cores at 1GHZ (opposed to 6 Jaguar cores at 1.6GHZ with a 1.84tflops GPU). In fact Matt's comments (this and the "Switch is in the middle between PS360 and PS4/X1, but closer to last gen") are the only thing that makes me believe the digital foundry leak.

I remember blu saying that Wii U ports could potentially cause some headscratches due to the low clock of the A57 cores, but Zelda, Mario Kart and Fast seem to run much better than on Wii U despite that. I think it would be nice if we could get Shin'en to comment on how smooth was to port FAST on Switch, it likely wouldn't tell us the whole story but maybe we can extract some extra context from their answer.

 

[Skittzo0413]

Foxconn clocks at 16nm sounds great and plausible, but Im not quite ready believe Eurogamer got it wrong. 20nm at eurogamer clocks sounds just as plausible.

What will it take for us to know for sure? Another leak? Can we tell from a die shot?

That's a good question... I don't think know if the process size is something a standard die shot can determine, as there really aren't that many differences between 20nm and 16nm FF... Unless someone has a very, very good scanning electron microscope and a TX1 for comparison. Edit: A72s vs A57s might be easier to determine from a die shot though.

It might be easier for someone to actually hack the SoC to determine its maximum clock rates and power draw at those clock rates. That would tell us for sure if it's 16nm vs 20nm.

 

[frankie_baby]

I don't think so, Matt simply said that Switch wouldn't be as CPU limited as PS4, which, considering it's 157gflops undocked and 393-472 when docked, it's only obvious, even with 4 A57 cores at 1GHZ (opposed to 6 Jaguar cores at 1.6GHZ with a 1.84tflops GPU). In fact Matt's comments (this and the "Switch is in middle ground between PS360 and PS4/X1, but closer to last gen") are the only thing that makes believe the digital foundry leak.

I remember blu saying that Wii U ports could potentially cause some headscratches due to the low clock of the A57 cores, but Zelda, Mario Kart and Fast seem to run much better than on Wii U despite that. I think it would be nice if we could get Shin'en to comment on how smooth was to port FAST on Switch, it likely wouldn't tell us the whole story but maybe we can extract some extra context from their answer.

Well it's worth noting shinen took years to make FAST on Wii U but seem to have done the port fairly swiftly

 

[Hermii]

That's a good question... I don't think know if the process size is something a standard die shot can determine, as there really aren't that many differences between 20nm and 16nm FF... Unless someone has a very, very good scanning electron microscope and a TX1 for comparison. Edit: A72s vs A57s might be easier to determine from a die shot though.

It might be easier for someone to actually hack the SoC to determine its maximum clock rates and power draw at those clock rates. That would tell us for sure if it's 16nm vs 20nm.[

Nintendo got that one covered by not having a browser. This thing wont be easy to hack.

Edit: If we can determine if its A72 or A57, we will pretty much know the process node. A72 rules out 20nm.

 

[Skittzo0413]

Nintendo got that one covered by not having a browser. This thing wont be easy to hack.

Right off the bat I will admit I know absolutely nothing about hacking, but in theory if somewhere were to break open their unit and attach contacts to the right places, wouldn't they be able to communicate directly with the SoC? Maybe that's not really hacking per se, but I would imagine that could be possible. Maybe not easy at all.

Edit: If we can determine if its A72 or A57, we will pretty much know the process node. A72 rules out 20nm.

Yeah that would probably be how we determine it. I'm guessing an A72 core would look different from an A57 core in some way, so that should be possible with a die shot.

 

[z0m3le]

Right off the bat I will admit I know absolutely nothing about hacking, but in theory if somewhere were to break open their unit and attach contacts to the right places, wouldn't they be able to communicate directly with the SoC? Maybe that's not really hacking per se, but I would imagine that could be possible. Maybe not easy at all.



Yeah that would probably be how we determine it. I'm guessing an A72 core would look different from an A57 core in some way, so that should be possible with a die shot.

It's much simpler than all of that. The die size would tell you everything you needed to know. If it's smaller than the X1 chip inside Shield TV then it is 16nm, if it is ~20% bigger than Shield TV's then it is 16nm with an extra SM or embedded memory.

For it to be 20nm, it would pretty much be the same size, thanks to the thermal pictures we can rule out more than 2SM at this node and Eurogamer's clocks, but we can't rule out SRAM bulking up the size, so there is a small chance for that to make the die bigger.

 

[nightside]

Iirc the Foxconn rumor doesn't talk about SMs right?

 

[Lexxism]

921MHz. No one knows or claims to know the core configuration, though it's still likely to be 2SMs so 256 cores. So it would be 471 GFlops.

Damn. Not even 50% of XB1 GPU.

 

[Rodin]

Damn. Not even 50% of XB1 GPU.

Apples to oranges, not only because we don't know how much fp16 can boost performances of the Switch. But also because of this.

Either way, no, it doesn't get to 50% of Xbox One, and there's a good reason why.

 

[Donnie]

Damn. Not even 50% of XB1 GPU.

In 32bit precision yes (though with better efficiency real world 32bit precision would be closer than raw numbers suggest). In 16bit precision it would be 942gflops. Of course you can never get away with using fp16 for everything, but it can be used quite heavily on Switch if a developer decides to do so. Same can't be said for XBox One.

 

[Skittzo0413]

It's much simpler than all of that. The die size would tell you everything you needed to know. If it's smaller than the X1 chip inside Shield TV then it is 16nm, if it is ~20% bigger than Shield TV's then it is 16nm with an extra SM or embedded memory.

For it to be 20nm, it would pretty much be the same size, thanks to the thermal pictures we can rule out more than 2SM at this node and Eurogamer's clocks, but we can't rule out SRAM bulking up the size, so there is a small chance for that to make the die bigger.

I don't think the thermal pictures rule anything out, as we don't know how long the unit was running when those pictures were taken, though obviously >2SMs seems unlikely. But would the die size really tell us if it's 16nm? I thought there was no change in density between 20nm and 16nm?

Iirc the Foxconn rumor doesn't talk about SMs right?

No but I wouldn't expect more than 2. 3 or more would likely increase the die size, and the leaker does state that it's 10mm x 10mm.

Damn. Not even 50% of XB1 GPU.

When you account for the architectural advantage of Maxwell/Pascal over the GCN in PS4/XB1 and 2x processing for FP16 it's probably closer to 75%.

 

[z0m3le]

In double precision yes (though with better efficiency real world double precision would be closer than raw numbers suggest). In single precision it would be 942gflops.

so, it's single precision for fp32 for 393 to 472gflops
double precision is fp64 (not used for gaming)
half precision is fp16 786 to 944gflops

Mixed precision estimate, based on an ex Ubisoft's developer's work: ~800gflops with Foxconn's clocks and ~670gflops with eurogamer's clocks

It's roughly half the performance of XB1 or half the performance of PS4, especially given Maxwell/Pascal vs GCN.

 

[Rodin]

When you account for the architectural advantage of Maxwell/Pascal over the GCN in PS4/XB1 and 2x processing for FP16 it's probably closer to 75%.

That honestly doesn't sound realistic to me. I believe FP16 does boost performances, but i doubt it can be used to the point that it gets the Switch to 75% of Xbox One in real world performances.

Mixed precision estimate, based on an ex Ubisoft's developer's work: ~800gflops with Foxconn's clocks and ~670gflops with eurogamer's clocks

It's roughly half the performance of XB1 or half the performance of PS4, especially given Maxwell/Pascal vs GCN.

I think you're referring to sebbbi here, do you have a link for this?

 

[Lexxism]

Apples to oranges, not only because we don't know how much fp16 can boost performances of the Switch. But also because of this.

Either way, no, it doesn't get to 50% of Xbox One, and there's a good reason why.

In 32bit precision yes (though with better efficiency real world 32bit precision would be closer than raw numbers suggest). In 16bit precision it would be 942gflops. Of course you can never get away with using fp16 for everything, but it can be used quite heavily on Switch if a developer decides to do so. Same can't be said for XBox One.

When you account for the architectural advantage of Maxwell/Pascal over the GCN in PS4/XB1 and 2x processing for FP16 it's probably closer to 75%.

Ah. So it might be 75% at best? That would be not quite bad compared with my initial post.

 

[frankie_baby]

With regards to fp16 i remember reading that UE4 supports using it natively, does unity?

 

[Skittzo0413]

That honestly doesn't sound realistic to me. I believe FP16 does boost performances, but i doubt it can be used to the point that it gets the Switch to 75% of Xbox One in real world performances.

Ah. So it might be 75% at best? That would be not quite bad compared with my initial post.

I don't know if I'd say 75% exactly, but I think it would likely get higher than 50% and possibly closer to 75% when you account for FP16 and improved architecture. Remember, Maxwell itself is even more modern than what's in the PS4 and XB1, and we can see on PCs how big of a difference that makes, comparing Maxwell Nvidia flops to first/second generation GCN flops.

I doubt the difference will be as pronounced on consoles as it is on PC, but that doesn't mean the difference won't be there. And then I recall hearing some pretty encouraging things about FP16 from Matt. I don't remember exactly what he said but I think he responded to someone who said something like "you can't use FP16 for any appreciable amount of code in a video game" by saying that's very wrong.

So I really do expect this to punch pretty far above its weight. Multiplats like Steep, Rime and Yooka-Laylee will be a good first test.

EDIT: Also the PS4 Pro getting 2x FP16 support will encourage developers to make more use of it, which will help immensely in getting these developers more situated with developing and optimizing for Switch.

 

[z0m3le]

https://www.semiwiki.com/forum/content/1789-16nm-finfet-versus-20nm-planar.html
My understanding is that it should be slightly smaller, something we might be able to tell from a die shot.

That honestly doesn't sound realistic to me. I believe FP16 does boost performances, but i doubt it can be used to the point that it gets the Switch to 75% of Xbox One in real world performances.



I think you're referring to sebbbi here, do you have a link for this?

Yes, him. It's buried in an 80+ page thread about the Switch, couldn't find it just now but probably somewhere between page 40 and 60.

 

[sits]

Out of curiosity, how do the CPU's compare? For example, we know the PS4 -> Pro CPU upgrade was about 1.3x, and we also know the X1 CPU is a fraction more capable than the Base PS4's CPU.

I'm just wondering because, if the Switch's CPU is not too drastically far behind the Base PS4, then maybe more third-party games might find their way on to the Switch, albeit running at a lesser resolution, maybe 720p/900p....?

....or maybe I'm just talking out of my arse.

 

[frankie_baby]

Out of curiosity, how do the CPU's compare? For example, we know the PS4 -> Pro CPU upgrade was about 1.3x, and we also know the X1 CPU is a fraction more capable than the Base PS4's CPU.

I'm just wondering because, if the Switch's CPU is not too drastically far behind the Base PS4, then maybe more third-party games might find their way on to the Switch, albeit running at a lesser resolution, maybe 720p/900p....?

....or maybe I'm just talking out of my arse.

Eurogamer clocks - 50-75% of PS4
Foxconn "leak" clocks - 100-150%