NVIDIA Tegra Architecture

Let's ask "technical marketing"....

 

That would be a good question to ask NVIDIA and Microsoft!

 

So is there a partial precision (_PP) flag that can be set within the framework of ps_2_0? See 10 year old discussion (!) here: http://forum.beyond3d.com/showpost.php?p=83692&postcount=12

 

http://www.nvidia.com/object/white-papers.html

Nvidia released PDF white papers on Tegra 4/4i, i500 & Chimera.

 

Microsoft is not crystal clear on that topic but my understanding is that ps_2_0 requires FP24, as does ps_4_0_level_9_2. However ps_4_0_level_9_1 would work with FPwhatever.

D3D1x.x ps_4_0_level_9_x are slightly different than D3D9.x ps_2_x and shaders have actually to be compiled against both profiles but precision doesn't matter there.

 

No they're not!

 

http://forum.beyond3d.com/showpost.php?p=1613873&postcount=15

I asked this last year when nVidia's Tegra 3 Windows RT tablets were announced but never really got a good response at that time.

http://www.tomshardware.com/reviews/tegra-4-tegra-4i-gpu-architecture,3445-3.html

That appears to be correct. According to Tom's Hardware, nVidia is only claiming Direct3D 9_1 support for Tegra 4 where FP20 is sufficient. Tegra 4 does support a number of 9_3 level features, like instancing, which they can't currently expose in Windows given the definition of the feature levels. Apparently nVidia is trying to convince Microsoft to allow some type of exception/compromise to expose their cherry-picked feature set, which I thought was precisely what Microsoft wanted to prevent when they eliminated capability bits in DX10 and later instituted well-defined, cross-vendor baseline feature levels.

EDIT:
http://msdn.microsoft.com/en-us/library/windowsphone/develop/jj714085(v=vs.105).aspx

An interesting note is that while Windows RT mandates a minimum of Direct3D 9_1, Windows Phone 8 actually requires Direct3D 9_3 on the dot. That would presumably mean both Tegra 4 and 4i are shut out from the Windows Phone market, which depending on how that OS grows in the next year or two, may or may not be important.

 

I'm still a bit confused, because 9_1 is still (unless I'm mistaken) ps_2_0 and thus 24-bit. Time to write some shaders I think.

 

Who cares really? I mean since they claim themselves that the difference between FP20 and FP32 is negligable (I'd swear they didn't claim anything similar during the NV30/R300 timeframe ) and we're trying to find whether 4 bits up or 4 bits down. All you need to know for now is that FP20 is perfectly sufficient for fragment shading because they say so and when they'll finally go for USCs they'll re-invent the wheel for FP32.

Even that is besides the point. I wonder if they even themselves believe the bullshit they're touting. I think in Tegra2 the GPU block was roughly 1/5th of the entire SoC die estate. That was with 1 Vec4 FP20 PS ALU, 1 Vec4 FP32 VS ALU, 2 TMUs at 333MHz under 40nm.

http://www.anandtech.com/show/4144/...gra-2-review-the-first-dual-core-smartphone/3

Could I have a similar die shot from T4 for comparison?



Now what I am supposed to believe here is that going from 40 to 28nm was enough to not ONLY reduce the GPU block to 1/8th of the SoC die area estate, but at the same time units increased to 12 Vec4 FP20 PS ALUs, 6 Vec4 FP32 VS ALUs, 4 TMUs and to that at 672MHz. I'm all eyes if someone would like to give me even an educated guess how something like that can be accomplished since I must be missing something essential here.

 

I struggle to see where Nvidia went terribly wrong on this one. I doubt WinRT was very much on the radar when Tegra 3 was being defined and it's not as Angry Birds suffers major image quality loss when using a 20 or 24 bits shaders.

If Microsoft decided to use Tegra 3, it's clear that they didn't care very much either. And neither does the market: I don't think I read a single Surface RT review lamenting the lack of pixel precision.

Good point: what's worse? The inconsequential lack of pixel precision or the lack of stability? Which one would be most damaging to Microsoft's brand?

 

Exactly. If lack of pixel precision is terrible, you're in danger of running out of superlatives fast for the stuff that matters.

That also means that Tegra 4 should be fine wrt graphics for whatever Surface or Android product that comes next.

 

Discussion of Tegra's roadmap

As far as I can tell it's only nVidia that has been pushing for custom enhanced versions of mobile games to take advantage of faster hardware. Everybody else just codes for the lowest common denominator. What exactly do you want Microsoft and nVidia to do differently?

In general it looks like all these SoC manufacturers are trying to sell hardware that very few people really want or need. I always wonder why people care about GLBenchmark scores on a phone. What exactly can you do on a phone with a badass GPU?

PPI on mobile devices is already soaring well past PC monitors. All of this graphics performance is just going to be wasted on rendering detail that nobody can see. I don't really see the point to be honest. I would love someone to come up with a compelling use case for very fast SoC graphics but I don't see one right now.

 

Why do folks use public benchmarks? Note that question is use and NOT abuse

Play mobile games while on the move?

You're not going to place a smartphone in a 2 meters distance to play any sort of game are you? Extremely high PPI are extremely useful for text amongst others.

Just because you fail to see the point it doesn't mean that some folks aren't killing some time on the move watching videos, playing games or whatever else with their mobile device.

When you don't see a point from the get go how high are the chances you'd find any sort of compelling use for high graphics performance in SFF mobile devices anyway?

 

http://www.pocketgamer.co.uk/r/iPad/Need+for+Speed:+Most+Wanted/feature.asp?c=46679
http://www.pocketgamer.co.uk/r/iPad/Modern+Combat+4:+Zero+Hour/feature.asp?c=47563

Arguably, the reason nVidia has to ask/help developers make custom enhanced versions of mobile games for their SoC is because Apple doesn't need to ask for developers to do it for their PowerVR GPUs. Need for Speed Most Wanted and Modern Combat 4, both high-end mobile games from two different big developers, run with noticeably better graphics on the SGX543MP2 iPad Mini compared to the Tegra 3 Nexus 7 while still maintaining playable performance. iOS vs. Android market pressures may be the initiating reason why games tend to turn out better on iOS more than inherent PowerVR vs nVidia differences, but presumably PowerVR Android devices get the iOS graphics enhancements for free since it's already in the code. nVidia pushing a selection of enhanced games may be less an attempt to come out on top of competing GPUs, but more an attempt to keep up with the silent presence of PowerVR-optimized games.

Well putting aside games, nVidia is saying that it enables more fancy camera features.

 

I don't think you're following me Ail. I play a lot of games on my phone and use my Nexus 7 every day. I'm not questioning the need for GPUs. Just the need for faster ones - in response to the comment about pushing the envelope.

All of the things you mentioned - mobile gaming, higher definition text - don't require blockbuster graphics performance. What can we do on tomorrow's SoC's that we can't do today?

 

Angry Birds!

Because watching videos and whatever else they do on their mobile device stresses current generation GPUs?

I'm fine with GPU perf increasing, of course. But I totally get why Nvidia decided to favor CPU over GPU for Tegra 3. Being known for a PC GPU company doesn't mean that you shouldn't first get the performance fundamentals right. And for the vast majority of users, things are still CPU bound. This is no different than Intel with PCs: focus on CPU first, once that levels off, start increasing the GPU.

 

Ok then since now it's a bit clearer: something for CAD on the move and maybe a pinch of ray tracing capabilities on the move here and there?