Nintendo 3DS hardware thread

[Exophase]

According to this it's listed as FCRAM in official documentation: http://www.3dbrew.org/wiki/Hardware

What do you mean a product that's not theirs? It's DMP's booth. They're showing a product running their IP, that would be at least partly representative of what you get if you license their IP at TSMC regardless of whether or not that product was itself fabbed at TSMC. The VRAM technology used in 3DS is irrelevant to DMP's presentation.

 

[Airwave]

is the cpu clock speed known yet for sure? somene says it could be around 800 mhz -1 ghz, since the arm 11 mpcore family of cpu's starts from that frequency going upwards. Do you find this possibility to be plausible?

 

[Exophase]

The 268MHz number on the 3dbrew wiki is probably reliable. I've seen it (roughly) verified before it was posted, by an ex-employee of a game company that was working on 3DS games. Since they're claiming to have seen official documentation for these values they'd be more outright lying than just mistaken.

A frequency used in marketing material doesn't mean that it can't clock (potentially much) lower, a power sensitive design like ARM11 would never be so limited to high clock speeds. Especially not since we've seen plenty of low frequency ARM11s before ARM launched ARM11MP, and I don't see what they'd have changed in the design to raise the minimum frequency a whole bunch.

 

[Airwave]

Then, don't you find a single core* cpu clocked so low completely disappointing for a 2011 device? isn't a 268 mhz cpu going to be a bottleneck in things like IA, physics, number of enemies on screen etc.?

*(since the other seems to be used only for os)

 

[Exophase]

Then, don't you find a single core* cpu clocked so low completely disappointing for a 2011 device? isn't a 268 mhz cpu going to be a bottleneck in things like IA, physics, number of enemies on screen etc.?

*(since the other seems to be used only for os)

I find it completely par for the course for a Nintendo handheld. Compare GBA's 16.7MHz ARM7TDMI, released in 2001, with StrongARM processors in excess of 200MHz having been in mobiles for years prior. Same with Nintendo DS's 66MHz ARM9, vs the contemporary 624MHz PXA270 XScale, or PSP's 333MHz MIPS32 4k for that matter.

In other words, their CPUs have been consistently at least five years behind the high end for mobile devices. That they're using an ARM11 to begin with should be a strong hint.

 

[Airwave]

i really hope that they at least unlock the other cpu to be used by developers, because as it is, the 3ds seems to be very weak on the cpu side, and i fear that it will significantly hold back the platform. So far i have only played Re revelations, and for example there i can't see any particular physic interaction, and few enemies on screen.

 

[Exophase]

There was a rumor, about a year old, that Nintendo was making the second CPU accessible to games. So it's possible that this already happened.

 

[Airwave]

yeah seems like RE:re was the first game (at least the first third party game) to benefit from the extra core. However even with the second cpu unlocked, the overall cpu power is still pretty low compared to other mobile devices on the market right now, i wonder how will it impact on the overall performance of the console.

 

[Airwave]

regarding the fillrate of the gpu, there's something i don't understand. Offical datasheets for the pica states that it can draw 15,3 Mpolygon/s @ 200 mhz , and that it has a fillrate of 800 Mpixel/s at that frequency. Aen't these data incongruent to each other? a fillrate of 800 Mpixel/s should give you 25 Mpolygon/s , (800: 32 in case of 32 bit pixels) or 33,3 (in case of 24 bit), so where does this 15,3 figure @ 200 mhz comes from? am i doing my calculation wrong?

 

[Exophase]

Fillrate and triangle or polygon rate are two completely different and independent metrics. And both of them, especially the latter, are pretty ambiguous. So the numbers vendors provide aren't that meaningful. Usually the fillrate number will be a peak number of texels that can be sampled (used to be it was fragments that can be rendered, something that would usually correlate with ROPs these days), while the triangle or polygon rate will be limited by something like triangle setup. But could instead be limited by vertex shading or even culling.

There's usually no implication of how many fragments the triangle or polygon is on average, it could in fact be zero.

 

[Airwave]

i wondered, if the 268mhz clock speed is true, then 3ds fillrate should be 1072 Mpixel/s (268 X 4 pipelines) a pretty high value for a handheld. higher than both gamecube, and wii. And even considering it halves, (because of the need to redraw the image 2 times), a value of 536 Mpixel/s is still not bad imho. If my calculation is right (536/32) 3ds could still draw a peak of 16,7 Milions of triangles per second, 556.000 tris per frame @ 30 fps, for reference the original far cry pushed more or less that many triangles per frame on average if i remember well.

 

[Exophase]

Why are you still correlating triangle rate with fill rate? :/

Here's an extreme example - take the original DS. It's rated at approximately 30Mpixel/s and 120Kpolygons/s. That's a ratio of 250, which totally dwarfs your arbitrary ratios of 32 or 24. But it's also completely correct. The GPU processes one pixel/clock at 33.514MHz, but has a blank period of 23/263 lines so is only active 91.3% of the time, which gives about 30.6M active cyles. The polygon count, on the other hand, is hard limited to 2048 per frame (and frame rate is fixed at ~60Hz) because it has to bin them in a limited amount of memory. So the limit is 2048 * 59.8261Hz = 122524 polygons (triangles in fact, with quads it's less).

Ultimately fill rate and triangle count are at best upper bounds for what you can do, even if everyone was actually using the terms in any meaningful and consistent way. Rarely do real world games come anywhere close to claimed rates, especially triangle rates. And if all those pixels you're drawing having really simple texturing and lighting the game will look a lot worse than one you could do with a lot less.

 

[Airwave]

I once read that you could calculate peak polygons/s capacity of a gpu from the pixel fillrate by dividing it by 32 (not an arbitrary value, but refering to 32 bit pixel iirc), the use of this formula in wikipedia pages (for example the polygon performance of the first xbox is calculated like that in the wikipedia page dedicated to that console) convinced me that it was a valid metohod.But if that's not the case i'm only happy to understand that . So can we take the 15,3 Mpopygon/s @ 200 mhz as a beleivable data? that would mean that the 3ds could push as a theoretical peak around 20 milion of polygons per second (10 if we consider the 3d?). that would mean that the console has polygon capabilities around ps2 level (that if i remeber correctly peaked at the 175.000 polys per frame @ 60 fps of jak 3 or it was one of it's most polygonaly complex anyway)

 

[Exophase]

It's not so much a matter of whether or not the 15.3MPolygon/s number is accurate or not (although I'm sure there were some cases where claimed numbers were just flat out lies) but that without knowing anything about what it means we can't draw any useful conclusions from them.

Consider that you're comparing one company's peak claims against peak figures from actual games. The arguably more fair (but still completely meaningless) comparison would be between what DMP claims and what Sony claimed for PS2, which is of course a much higher number than the one you gave. We know all about how PS2's GS and EE work but we know very little about Pica200 so it's really hard to make a useful comparison, except that it's obvious 3DS can do effects per-pixel that PS2 can't (at least not directly).

 

[Airwave]

well, i'd say that without having accesss to the dev kits or the gpu itself, the very discussion of this thread would risk to be meaningless, because i think we are here trying to better understand the inner workings of the console, and making educated guesses of its perfromance, but i find this discussion interesting nonethless. Of course I'm not looking to have a 100% exact figure of 3ds polygon performance, as trying to figure out if it's "real-world" polygonal performance are nearer to say a dreamcast or a xbox, to have a reference, in order to know what to reasonably expect from the console.

 

[Arwin]

Bend suggested that with their Golden Abyss engine they could get over 200.000 polygons per frame on the Vita. It was the least of their bottlenecks, and they ended up layering polygonal detail over a lot of stuff where doing textures or whatever would have been harder (considering that both storage memory and texture capability is more limited)

It's still a pretty high number of polygons for a portable already, and that's at 30fps, not 60fps. Which when comparing to 3DS games, leads me to believe that the theoretical number of 255k polygons per frame for the 3DS is very theoretical (and because of 3D every frame has the 60fps budget, so divide by 2 again), and even then most of the bottlenecks are likely to be elsewhere ...

But we do know fairly sure that the 3DS is not faster at pushing polygons than the original Wii, as it is a rendering of the same environment as Wii Sports Resort, which you can fly through the same way, and it has had to reduce the detail in the polygons a fair bit from the Wii version. And stereo display is 30fps, so taking that out of the equation doesn't help.

 

[Grall]

I once read that you could calculate peak polygons/s capacity of a gpu from the pixel fillrate by dividing it by 32

That's not true. It's in fact nonsense. Polys/s and pixel fillrate are entirely different things and are not related at all. In fact, both metrics aren't actually metrics at all, and haven't been so for many years now.

Depending on what kind of triangle or pixel one is talking about a modern GPU would have wildly different performance. Typically only peak numbers are talked about, but no software uses flat-shaded, single-textured, rigid polygon models anymore. Games today uses pixel shaders and multitexturing for polygons, and vertex-shaded, soft-skinned models for characters and at times parts of environments too.

(not an arbitrary value, but refering to 32 bit pixel iirc)

That also does not make sense. Pixel bit depth is another, unrelated factor in this.

Taking fillrate and divide by 32 would only give the maximum number of 32-pixel polygons the GPU could draw - theoretically. But since polygon setup, culling, shading etc is handled by different hardware on the GPU than pixel fill, it's subject to different limitations. It could possibly draw this many polygons or more (probably more for modern hardware, as a 32-pixel poly is fairly large by today's standards, for desktop GPUs anyway. Mobile is a different matter entirely), or it might be less. Pixel fill and polygon rates are unrelated.

 

[Airwave]

So, considering a clockspeed of 268 mhz, we have to expect something in the 10 milion polys/second?, (maybe less if that data is a theoretical figure?) That would be less than i expected to be honest, games like monster hunter tri (that looks remarkably similar to the wii counterpart) led me to think that it was capable of gamecube/wii geometry, i guess The small screen probably helps in the perception of detail.

@ gral
thanks for claryfing that. So thre's no way we can tell polygon performance from other data. Seems like we'll have to take data revealed by manufacturers/programmersfto have an idea