Hardware Nintendo Switch Technical Specifications Discussion

[ned]

I honestly think for gaming to survive into the next generation of consoles the whole idea of
a gpu needs a massive rethink, paying artists to come up with realistic looking 4-16k
texture assets doesn't really seem feesible to me, I think we've already hit a wall with 1080p.
What's needed is a gpu that renders textures procedurally based on daily weather reports with
different timezones for all, where all that's need is for the developers to program a base colour for
each 3d object with custom shaders, even down to logos in a texture being totally raised from the surface
or strands of material fibres.

 

[TotalInsanity4]

I honestly think for gaming to survive into the next generation of consoles the whole idea of
a gpu needs a massive rethink, paying artists to come up with realistic looking 4-16k
texture assets doesn't really seem feesible to me, I think we've already hit a wall with 1080p.
What's needed is a gpu that renders textures procedurally based on daily weather reports with
different timezones for all, where all that's need is for the developers to program a base colour for
each 3d object with custom shaders, even down to logos in a texture being totally raised from the surface
or strands of material fibres.

That's entirely done by software, though. You may as well just make that the foundation of the game's rendering engine as opposed to baking it into the gpu

 

[ned]

That's entirely done by software, though. You may as well just make that the foundation of the game's rendering engine as opposed to baking it into the gpu

I'd like to see the terminator programming his own games so to speak, self learning AI has come on leads and bounds over the last few years.

 

[TotalInsanity4]

I'd like to see the terminator programming his own games so to speak, self learning AI has come on leads and bounds over the last few years.

I agree, but again, that's api/rendering engine tasks, not hardware

 

[Yil]

640 Cuda cores is probably the most likely, quite a nice number actually.

Do you think they stacked the two on top of each other or on two sides of a pcb which sounds ridiculus?
Or maybe Nintendo intentionally leaked false information to Eurogamer. I think Foxconn is still partially true (the original post in Chinese Reddit is dead, even if Reddit is still alive).

--------------------- MERGED ---------------------------

I agree, but again, that's api/rendering engine tasks, not hardware

But gpu can do more than drawing dots and calculation even with instruction sets (somehow gpu instructions are entirely ignored when graphic api slowly raises), can they not?

 

[Xzi]

I honestly think for gaming to survive into the next generation of consoles the whole idea of a gpu needs a massive rethink

Here's the fun thing: it already happened. With VR taking more mainstream prevalence, re-projection/timewarp methods have given way to the ability to effectively double any modern GPU's power. I have no doubt that some of that same technology is there on both a hardware and API-level for Switch. People will be more impressed than they expect to be, especially for a device at ~$250.

This also gives rise to another issue, though: this is pretty much it for perceivable upgrades in graphics on a technical level. Switch's performance can be effectively doubled, but so can my GTX 1070 or someone's Titan. From here games will have to decide whether to rely on an aesthetic choice, go for full-blown realism, or just trust in your artists to make it great. I think as-is typical, people will shun samey graphical styles regardless of the resolution or detail and be attracted to what stands out.

Side note: trying for realism doesn't often work out, but Christ that RE7 demo in 4K is fantastic.

 

[TheDarkGreninja]

Do you think they stacked the two on top of each other or on two sides of a pcb which sounds ridiculus?
Or maybe Nintendo intentionally leaked false information to Eurogamer. I think Foxconn is still partially true (the original post in Chinese Reddit is dead, even if Reddit is still alive).

--------------------- MERGED ---------------------------


But gpu can do more than drawing dots and calculation even with instruction sets (somehow gpu instructions are entirely ignored when graphic api slowly raises), can they not?

Just thought about that, guess it pretty unlikely after reading up the consoles dimensions.

--------------------- MERGED ---------------------------

So it seems the 71 people who said pascal were wrong, including me. The tegra X1 is an 8-core right? If only this machine was using A73's...

 

[Foxi4]

Maxwell with lowered frequency, but higher core count makes absolute sense with wattage, heat dissipation and cost considerations in mind. I've been saying that Pascal sounds too good to be true for a while now, especially prior to Hate Train #1. I'm sure @TheDarkGreninja remembers our convo, but for the article I followed the prevailing theory with a dose of skepticism. I'm not surprised that Nintendo downscaled the construction - they had to. 1 teraflop is respectable when docked, the bottleneck here is undocked mode as I assume that games need to work in both modes, and since you get more or less half the clock when undocked, the 500 gflops estimate was accurate. I'm glad to see cooling vents, which are likely active, enabling the system to go beyond the mark of passive cooling when going full-blast - that's what enables the system to dial the clock up when docked. The fan probably kicks in when the system is connected, this makes sense. Still substantially weaker than the PS4 or even the XBO, but for a portable console it's respectable.

 

[Yil]

Maxwell with lowered frequency, but higher core count makes absolute sense with wattage, heat dissipation and cost considerations in mind. I've been saying that Pascal sounds too good to be true for a while now, especially prior to Hate Train #1 - I'm sure DarkGreninja remembers our convo, but for the article I followed the prevailing theory with a dose of skepticism. I'm not surprised that Nintendo downscaled the construction - they had to. 1 teraflop is respectable when docked, the bottleneck here is undocked mode as I assume that games need to work in both modes, and since you get more or less half the clock when undocked, the 500 gflops estimate was accurate. I'm glad to see cooling vents, which are likely active, enabling the system to go beyond the mark of passive cooling when going full-blast - that's what enables the system to go full speed when docked. The fan probably kicks in when the system is connected, this makes sense. Still substantially weaker than the PS4 or even the XBO, but for a portable console it's respectable.

Do you think someday the only surviving member of gaming is steam/ gog and Nintendo?

 

[Foxi4]

Do you think someday the only surviving member of gaming is steam/ gog and Nintendo?

I think Nintendo will eventually become a full-time software developer as console gaming integrates with our smart devices, just like casual cameras, portable media players, PDA's and PNA's before it. Sony and Microsoft are trying to avoid this by establishing ecosystems - PlayStation 4 and Xbox One might be the last two systems in the lines, from this point onwards it could just be PlayStation and Xbox, turned into services running on set-top "smart" entertainment boxes that get occasional upgrades. I don't think Nintendo is ready for such a switch, pun not intended, so I predict that they will fold eventually, first with home consoles, then with portable ones. Nintendo's value is in their catalogue of IP's, not their hardware.

 

[nmkd]

Do you think someday the only surviving member of gaming is steam/ gog and Nintendo?

This is my dream - only PC and Switch.

Not because I hate PS4/XB1, but because I want all my friends to be on the same platform that I am.

//Sent from my glorious OnePlus X

 

[Sizednochi]

At 768MHz Maxwell you're looking at about half the power of an overclocked GTX 970, or roughly the same power as a GTX 950

Y'know, a GTX950 is marginally stronger than PS4's GPU. I don't think that's right.

 

[Foxi4]

Y'know, a GTX950 is marginally stronger than PS4's GPU. I don't think that's right.

Just saying "Maxwell" is not enough - that's the architecture. We need to know the core count, not just the frequency, so you're right. Fortunately we have a rumoured flops count, so we can make some calculations and figure things out ourselves - some people in the thread already did just that.

 

[Xzi]

Just saying "Maxwell" is not enough, that's the architecture, we need to know the core count, not just the frequency, so you're right. Fortunately we have a rumoured flops count, so we can make some calculations and figure things out ourselves - some people in the thread already did just that.

You said 1TFlop in a previous post, yeah? Did that come from the Foxconn leak? Eurogamer is saying it would be about 567GFlops docked with 3 SMs. Possible then that they're getting the extra performance from some form of async re-projection? Would put it right at 1.1Flops v XB1's 1.3TFlops if you double the 567.

 

[Sizednochi]

Yeah, I think it'll actually be pretty close to Xbone One's GPU, if not on-par. Docked, that is.

Undocked, I wonder if the switch to 720p will make it up for the power difference.

 

[Foxi4]

You said 1TFlop in a previous post, yeah? Did that come from the Foxconn leak? Eurogamer is saying it would be about 567GFlops docked with 3 SMs. Possible then that they're getting the extra performance from some form of async re-projection? Would put it right at 1.1Flops v XB1's 1.3TFlops.

That's the figure I read today, the source was actually VentureBeat, I think. I don't think that Nintendo is using any form of trickery, they like things simple, if anything, people are mixing up different measurements. This actually *is* a source for concern, as nobody specifies "what kind of flops" we're talking about, FP16 or FP32. Here's for hoping that it's FP32, otherwise it's not much beyond a bog standard X1. I can only hope that it is an X1 with double the GPU cores, and I think that's what I'll stick with as far as predictions go, based on what we know.

 

[Yil]

You said 1TFlop in a previous post, yeah? Did that come from the Foxconn leak? Eurogamer is saying it would be about 567GFlops docked with 3 SMs. Possible then that they're getting the extra performance from some form of async re-projection? Would put it right at 1.1Flops v XB1's 1.3TFlops if you double the 567.

Eurogamer never said anything about teraflop.
Also Efficiency is no longer linear with performance. You can have half the performance with 1/8 of power on an amd gpu of that die size if Amd bothered, I think it's the same case, only not as dramatic for Nvidia. On the other hand this makes overclocking AMD cards harder than Nvidia due to higher power draw.

 

[Xzi]

Hmm okay, looked up the VentureBeat article and they're definitely referencing FP32 TFlops, they mention PS4 at 1.8. Eurogamer is saying FP32 it's only at 567GFlops (again assuming 3 SMs), so a difference by half on these claims.

 

[Foxi4]

Eurogamer never said anything about teraflop.
Also Efficiency is no longer linear with performance. You can have half the performance with 1/8 of power on an amd gpu of that die size if Amd bothered, I think it's the same case, only not as dramatic for Nvidia. On the other hand this makes overclocking AMD cards harder than Nvidia due to higher power draw.

In-game performance will become increasingly reliant on the number-crunching capabilities of the GPU in the next few years due to reliance on GPGPU, PhysX, CUDA, Mantle etc., but you're right - synthetic benchmarks and theoretical values vary greatly from real-life implementations.

Hmm okay, looked up the VentureBeat article and they're definitely referencing FP32 TFlops, they mention PS4 at 1.8. Eurogamer is saying FP32 it's only at 567GFlops (again assuming 3 SMs), so a difference by half on these claims.

That's the thing - many reports just cover the numbers without the fine print, when in reality the difference between FP16 and FP32 is paramount. They just compare number X (PS4 tflops result) with number Y (Switch's rumoured performance) without checking if they're not comparing apples and oranges. I already covered this - it's the Nvidia Tegra scam - they boast about FP16 numbers when they don't mean squat, that's the source of all the optimistic articles on Nintendo-oriented blogs ran by people who have no idea what they're writing about. Nobody does the math, so you have different sources with different figures for the same piece of silicon.

 

[Xzi]

That's the thing - many reports just cover the numbers without the fine print, when in reality the difference between FP16 and FP32 is paramount. They just compare number X (PS4 tflops result) with number Y (Switch's rumoured performance) without checking if they're not comparing apples and oranges. I already covered this - it's the Nvidia Tegra scam - they boast FP16 numbers when they don't mean squat, that's the source of all the optimistic articles on Nintendo-oriented blogs ran by people who have no idea what they're writing about.

Hard to tell which article got it right if either, then.