Nintendo Switch Tech Speculation discussion

[Goodtwin]

Where it works, yes. Just not everywhere, as you suggested when you claimed Maxwell's performance-per-GFLOPs would be 2x GCN 2's.

Sorry, I wasn't clear. I kind of lumped the efficiency advantage of Maxwell in with its ability to use FP16 shaders. Maxwell vs GCN isn't anywhere near 2x the work per flop. I just think it's important to keep in mind that upwards of 70 percent of shaders can use FP16, and that does offer a 2x boost compared to the same work being done with FP32. I'm sure there is some inefficiencies so it's probably not quite 2x, but better than simply judging it by its FP32 thoughput.

Sent from my SM-G360V using Tapatalk

 

[Shifty Geezer]

Very interesting that Ubisoft has already been using checkerboard rendering for a while now. I would have to think that this would be very beneficial for developers trying to port AAA games to Switch.

Surely it's the opposite?! If they weren't using checkerboard on consoles already, that'd be an option for Switch to close the gap a little. As it is, the consoles are already being pushed hard enough that it's better to render half the pixels, and that'll be the same case with a down-port.

 

[Goodtwin]

Surely it's the opposite?! If they weren't using checkerboard on consoles already, that'd be an option for Switch to close the gap a little. As it is, the consoles are already being pushed hard enough that it's better to render half the pixels, and that'll be the same case with a down-port.

If they are rendering at 1080p but with checker board rendering, then rendering that same scene in 720p with checkerboard rendering is still rendering a lot less pixels, thus freeing up performance.

 

[CaffeinatedChris]

Surely it's the opposite?! If they weren't using checkerboard on consoles already, that'd be an option for Switch to close the gap a little. As it is, the consoles are already being pushed hard enough that it's better to render half the pixels, and that'll be the same case with a down-port.

From my perspective it's moreso that they've established a method of checkerboard rendering that provides a tangible boost in performance without resulting in end-user pushback due to image quality compromise.

They've already figured out how to squeeze blood from a stone - the question is, does that scale down to pebbles?

 

[Shifty Geezer]

If they are rendering at 1080p but with checker board rendering, then rendering that same scene in 720p with checkerboard rendering is still rendering a lot less pixels, thus freeing up performance.

Sure. My point is the relative performance is no better for Switch, so AAA ports won't benefit any more than otherwise. You were implying that checkerboard rendering would help close the gap a little, at least as I interpret what you wrote.

In real terms, checkerboard rendering could help Switch take on ports of games that don't use reconstruction* on consoles.

* Really must stop using 'checkerboard' as a generic catch-all term!

 

[Goodtwin]

Sure. My point is the relative performance is no better for Switch, so AAA ports won't benefit any more than otherwise. You were implying that checkerboard rendering would help close the gap a little, at least as I interpret what you wrote.

In real terms, checkerboard rendering could help Switch take on ports of games that don't use reconstruction* on consoles.

* Really must stop using 'checkerboard' as a generic catch-all term!

Honestly I was surprised to hear that so many games were already implementing the technique, but what I was implying was that the technique is available to use even at lower resolutions, such as 720p. I first started hear a lot about it with the PS4 Pro using it to render at 4K. Its all about reducing pixels to render. For example, if said developer were tasked with porting a native 1080p game to Switch, and rendered at 720p but with checkerboard rendering, this is a reduction from 2,073,600 pixels to just 460,800 pixels rendered.

 

[Blazkowicz]

Isnt the PS4/X1 GPU based on the GCN 1.1 architecture?

AMD called the RX480 and siblings "GCN 4", so the GCN 1.0, 1.1 and 1.2 (unofficial names made by the public) are ret-conned into GCN 1, GCN 2 and GCN 3.

 

[Goodtwin]

AMD called the RX480 and siblings "GCN 4", so the GCN 1.0, 1.1 and 1.2 (unofficial names made by the public) are ret-conned into GCN 1, GCN 2 and GCN 3.

Ok, but isn't the PS4 GPU basically an HD7790? I was under the impression that we are a couple generations newer with the current RX series GPU's.

 

[Blazkowicz]

Yes basically. In other words a HD 7790 is GCN2 (as long with R9 290 and 390), and that's what TottenTranz indirectly refers to when talking of GCN2-ish. That's what I attempted at making explicit

 

[Goodtwin]

Yes basically. In other words a HD 7790 is GCN2 (as long with R9 290 and 390), and that's what TottenTranz indirectly refers to when talking of GCN2-ish. That's what I attempted at making explicit

Ok, so Maxwell still outperforms GCN2 in pretty much every way outside of GPU Compute.

Sent from my SM-G360V using Tapatalk

 

[AlNom]

Ok, so Maxwell still outperforms GCN2 in pretty much every way outside of GPU Compute.

This is a weird comparison statement.

 

[Goodtwin]

This is a weird comparison statement.

Why? Does it not? I'm not saying it's a night and day difference, probably very modest.

Sent from my SM-G360V using Tapatalk

 

[Blazkowicz]

Ok, so Maxwell still outperforms GCN2 in pretty much every way outside of GPU Compute.

Sent from my SM-G360V using Tapatalk

But GCN2 (like other versions) has strengthes that came to use in games, a few years after 7970 and 7790 etc. were launched. New drivers, new games and eventually new APIs (that I will even enumerate : Xbox One, PS4, Mantle, DX12, Vulkan. I just threw the Xbox in the list to keep it simple for me). Async compute, e.g. some form of mutl-tasking where you can run Graphics, Compute and Data Transfer tasks somewhat in parallel, this allows GCN2 to more fully use up the chip's flops by e.g. combining a flops heavy tasks with a bandwith heavy task so that's faster than doing one after another.
Seems to make it catch up decently (if not quite) with Nvidia's. The GPU Compute lead you hint to was brought to use into actual games, this is what I'm trying to say.

But Maxwell is more recent, and more notable may be it has memory bandwith savings than GCN2 lacks.
Hereby I believe the Swith's Tegra X1 is a plain one and will just be bandwith starved. It'll be a poor thing with a single 25.6GB/s memory pool. But it'll be better than a similar chip with four A57, GCN2 and 25.6GB/s, if that were to exist.

 

[AlNom]

More that I don't know what you are quantifying ?

 

[Syferz]

So new shield is still X1, I can't imagine them sticking to 20nm, probably moved it down to 16nm, switch moving with the same chip makes sense. Would explain the longer battery life from final devkits, wonder if they changed clocks with the final hardware, since the rumor is October devkits are more powerful.

 

[Esrever]

new shield is x1 is 20nm

 

[cheapchips]

I'm imagining at this point that nVidia has an x1 stockpile so big it's warehouse is used to hide the Ark of the Covenant.

 

[Rootax]

They only changed the pad for the "new shield"... And I'm ok with that, the existing Shield is already powerfull enough pour what it does.

 

[Deleted member 13524]

Ok, but isn't the PS4 GPU basically an HD7790?

In architecture, yes. But the PS4's Liverpool is sub more powerful than a full Bonaire in certain areas, such as fillrate.

The point I was addressing was whether the dev tools were optimized for GCN.
Even if Switch were using GCN, Sony and Microsoft would have more incentive to not help Nintendo optimize its tools.

You're correct, but I think what @Rootax meant by dev tools was not APIs but rendering engines and techniques.
In 2017, multiplatform developers will have to make games for 4 consoles in established ecosystems all using a very compute-centric architecture from the same IHV.
Within this universe, taking advantage of Maxwell's perks (or going around its disadvantages) will be harder.

 

[Goodtwin]

I'm imagining at this point that nVidia has an x1 stockpile so big it's warehouse is used to hide the Ark of the Covenant.

I know your joking, but it does make you wonder if there is a little truth to this statement. Who knows just how optimistic Nvidia was with their sales expectations for the Shield TV. Has anyone seen sales figures for the Shield TV? Maybe Nvidia expected to sell 20 million Shield TV's and produced enough Tegra X1 Processors. Who knows, there is an argument to be made that its possible the Switch uses the stock Tegra X1, and maybe Nvidia has insane stock piles that they are selling to Nintendo for very cheap.

Seeing as how Nvidia says custom Tegra, I am holding out hope that its not a stock Tegra X1, but modified and improved for the sole purpose of a gaming machine. There has been a lot of speculation that the dev kits have at least up until October nothing more than the Jetson TX1 kit that is available for anyone to buy. Who knows, perhaps all but Nintendo's closest partners may not receive final dev kits until after the blow out next week in order to keep the specs secret. Just depends on how proud they are of this custom Tegra. Its purse speculation on my part, but using the Jetson TX1 as a dev kit is perfectly acceptable, even this late in development, if the final unit actually outperforms the Jetson X1 dev kits. Something like Memory bandwidth being improved with the final hardware isn't a problem. Giving developers a higher performing dev kit to finish up development in the last couple months before launch would only make finalizing development easier. Assuming the 256 Cuda cores are stalling thanks to the minimal 25GB/s memory bandwidth, how much would you really expect a games performance to increase by doubling it to 50GB/s? Maybe a 20-30% increase in framerate at the most?