Nintendo 3DS hardware thread

brain_stew

Regular
Now the 3DS is confirmed, we can leave behind the rumour mongery and focus on facts as they are revealed. This post will be updated with confirmed specs as they become known.

GPU : DMP PICA200 (press release)


___

Nintendo just confirmed that its using DMP's PICA200 technology, as was rumoured earlier in the thread.

There's some slides (in English) about the demo Futuremark put together for them that I found here:

http://journal.mycom.co.jp/photo/articles/2006/08/15/siggraph07/images/003l.jpg

Taken from this japanese article:

http://journal.mycom.co.jp/articles/2006/08/15/siggraph07/index.html

I hosted the video of that tech demo on Youtube a few days ago as it took forever to download from DMP's site. Video is here:

http://www.youtube.com/watch?v=3A-xxUyJvQQ

In summary, its an OpenGL ES 1.1 part with a whole raft of custom extensions. Vertex shader but no programmable pixel shader, however the fixed function pipeline can pull off many of the effects you see in OpenGL ES 2.0 games anyway.

No word on which specific model Nintendo went with at this point.
 
I don't really mind the fixed function pixel pipes ... but the lack of AA ... bleh.

Quaz51 seems to think that Nintendo are implementing 2x supersampling through jittered samples when you switch the screen to 2D mode, rather than increasing the resolution as many speculated. So if you're happy to give up stereo3D you can have your nice image quality.

His post is here:

http://www.neogaf.com/forum/showpost.php?p=21973337&postcount=695
 
I don't really mind the fixed function pixel pipes ... but the lack of AA ... bleh.
AA is supported through hardware.
aoMZq.png
 
I think this is the progression:

SGX530 - 1x TMU, 2x USSE,
SGX535 - 2x TMU, 2x USSE, doubles triangle setup rate over 530
SGX540 - 2x TMU, 4x USSE
SGX543 - 2x TMU, 4x USSE2, doubles depth/stencil (hidden surface removal rate) over 540, does some other stuff to improve bandwidth and increases triangle setup rate another 50%

The earlier cores were part of Series5, and 543 is part of Series5XT, so I think they consider it a bigger core change despite only increasing the product number by 3 (maybe because the number of TMUs/ALUs stays the same)
With the revelation of the PICA200, it's not as relevant anymore, but how does the SGX545 compare to the SGX540 and SGX543?

In terms of the PICA200 and 3DS, does that mean Nintendo will be using OpenGL ES, ie. it should be relatively easy for developers to port their existing iPhone/Android/etc. games to the 3DS? A comparable catalogue of cheaper games could certainly level the playing field with the App Store, while still being a platform for dedicated gaming.
 
Guess we're back to speculating how 3DS is accomplishing backwards compatibility, or at least I am. I imagine Nintendo considers it so mundane as to not be worthy of showcasing, but I do wonder if it's 100% and without caveat.

It feels to me that any mobile technology that doesn't boast "tiling", "deferred rendering/shading", or at least "early z" is kind of a disappointment. Of course, it could have any of these things, but I'm sure it'd be more willing to boast if it did. Just being an OGL ES 1.1 part makes it seem like it's pretty old, although 4 pixel/clock is pretty good (but then, PSP does that too) - probably necessary just to keep up with overdraw.

The anti-aliasing must come at a price, or else everything would be using it.. and so far just about nothing is.

I'm by no means as educated on this as most people here, especially in terms of applications, so could someone tell me exactly what exact features they think the fixed function "Maestro" features correspond with in implementation? Is "refraction mapping" just normal mapping, or environment mapping? Is per-pixel lighting dot3 shading? Is subdivision primitive tessellation? What would shadow include beyond what the stenciling can already do - maybe polygon ID features like in DS? Finally, gaseous object rendering o_O I assume that's not just fogging or something comparatively simple?

There are some siggraph slides on their procedural textures, but only with picture examples, not explanations. These look like some kind of somewhat random gradients:

http://journal.mycom.co.jp/photo/articles/2006/08/15/siggraph07/images/025l.jpg
http://journal.mycom.co.jp/photo/articles/2006/08/15/siggraph07/images/026l.jpg

While these just look like 8-way mirroring:
http://journal.mycom.co.jp/photo/articles/2006/08/15/siggraph07/images/027l.jpg

Further slide:
http://journal.mycom.co.jp/photo/articles/2006/08/15/siggraph07/images/029l.jpg

Seems to be addressing subdivsion, not sure in which direction the work is being done:
http://journal.mycom.co.jp/photo/articles/2006/08/15/siggraph07/images/032l.jpg
http://journal.mycom.co.jp/photo/articles/2006/08/15/siggraph07/images/033l.jpg
 
AA is supported through hardware.
When I hear FSAA I think supersampling. Even if it's multisampling there is no mention of Z-buffer compression.

If it costs 50% or more performance in fillrate limited circumstances then it probably won't be turned on very often.
 
Just speculating here, but the per-pixel feature set reeks of EMBM and DOT3. The "procedural texture" examples can all be recreated with/probably are perturbed 1D textures.

Subdivision could be n-patches ala Ati's TruForm, if anybody remembers that. Basically, the caveat of your wonderful organic round objects lies in how much work you have to do to stitch them together with artificial, hard-edged objects without getting creases and cracks and t junctions and whatnot. The technique also requires per-vertex normals, which probably don't even exist in the source data of most DS game 3D models, so that's an extra expense.

The only thing that might actually be new is the "per-vertex subsurface scattering" stuff, which is just a weird claim to make. It could be a non-linear color interpolator. Just as well it might be a byproduct of the tesselation engine.

I don't see tesselation or whatever the sss tech turns out to be getting any use at all in games. This stuff just isn't compatible with the modeling tools on the market, and honestly not worth the investment to change the tools around to match.

DOT3 per pixel is useful though, to make dynamic highlights that don't swim around between vertices, and we'll see EMBM used for water effects and the occasional bumpy reflection on a shield or something.

edit: would like to know more about their shadowing features. Standard shadow maps, or proprietary mojo?
 
Last edited by a moderator:
My take is that a Japanese company (Nintendo) wanted to buy from another Japanese company, and the end result is we get a very lackluster GPU in 3DS.
 
Jon Peddie seems to disagree regarding that "lackluster" thing.
He doesn't make a strong case about it though. Fixed function blocks and the ability to render fuzzy volume objects aren't all that exciting. No idea why he chooses to call that "amazing high-end graphics functions".

800MPix/s is the actual big news here ... assuming the 3DS uses the full-on version, and not some cut-down variant.
 
He doesn't make a strong case about it though. Fixed function blocks and the ability to render fuzzy volume objects aren't all that exciting. No idea why he chooses to call that "amazing high-end graphics functions".
He seems somewhat excited about the approach to use parametric engines instead of shader units (at least that's the main aspect he highlights). Hard to tell what exactly can be done with that. I guess it's mostly about efficiency/ power consumption.

800MPix/s is the actual big news here ... assuming the 3DS uses the full-on version, and not some cut-down variant.
The specsheet posted in this thread is outdated (issued 2006). There's a 2008 version, claiming the same fillrate but 40 million polygons per second at 100MHz, a clockspeed of up to 400MHz, and a power consumption of 0.5 - 1 mW per MHz at 65nm.
 
Last edited by a moderator:
He doesn't make a strong case about it though. Fixed function blocks and the ability to render fuzzy volume objects aren't all that exciting. No idea why he chooses to call that "amazing high-end graphics functions".

800MPix/s is the actual big news here ... assuming the 3DS uses the full-on version, and not some cut-down variant.

It is not that high fill-rate wise. PSP already pushed 664 MPixels/s targeting 480x272 resolution-wise (textured). We are still at 4 pixels per clock.
 
Here's that 2008 slide with the much more appealing numbers:

ws4d3l.jpg


Taken from here.

Pretty impressive that they managed to hit 400mhz at 65nm and those power consumption figures are very encouraging (and likely the major reason why Nintendo chose the platform). I wonder what Nintendo have managed to eek out of this thing at 40/45nm?

I figure these figures must apply to the "full fat" version with 4 pipelines and 4 vertex shaders but we have no indication whether this is the variant Nintendo is using.
 
Here's that 2008 slide with the much more appealing numbers:

ws4d3l.jpg


Taken from here.

Pretty impressive that they managed to hit 400mhz at 65nm and those power consumption figures are very encouraging (and likely the major reason why Nintendo chose the platform). I wonder what Nintendo have managed to eek out of this thing at 40/45nm?

I figure these figures must apply to the "full fat" version with 4 pipelines and 4 vertex shaders but we have no indication whether this is the variant Nintendo is using.

I am kind of wondering what the object rendered on the bottom right of that picture is... well, I have an idea... but it might be well... it would have to be brown in that case :p.
 
Back
Top