Hardware [Question] Nintendo 3DS CPU FLOPs?

[CrazySquid]

Hi there, I'm curious about some 3DS specs. I think this has been debated enough in the past, but I didn't find any particular numbers on the 3DS CPU actual floating point capabilities.


I know the following about the 3DS specs (New):
ARM11 CPU: 4x MPCore/4x (1?) VFPv2 Co-Processor at 804MHz. Includes 2MB L2 cache (1.25 DMIPS/MHz, but since it has L2 cache let's bump that to 1.4 DMIPS/MHz as a slight boost = 4502 DMIPS)
RAM: 256MB FCRAM (Old 3DS was 3.46GB/s, maybe New 3DS has faster RAM since it's a different chip?)
VRAM: 6MB (I saw at the documents, and actually New 3DS has 6MB of VRAM and not 10MB as Digital Foundry reported, can't find info regarding bandwidth but it's likely higher than main memory)
GPU: PICA200 at 268MHz (6.4 GFLOPS, 20.5 Million polygons per second at 268MHz, calcutated from 15.3 Million figure and 4.8 GFLOPS at 200MHz)

Now, the question would be how much flops per cycle can the CPU do? I'm reading this, and IF I'm not mistaken, the CPU can do 4 instructions per clock, since it's 2 DP MUL MAC, which translanslates into 4 FLOPs.

So... with that, the CPU performance would be: 804MHz (CPU speed) x 4 (Single VFPv2 FLOPs) x 4 (VFPv2 number of Co-Processors) = 12.8 GFLOPS.

However, according to Nintendo, the CPU specs are the following:

As you can see, they don't mention 2x VFPv2, but rather, just a VFP co-processor for the CPU (I assume, the whole processor and not a VFP per core). Also, there's no mention of adding new VFPv2 processors into SNAKE (New 3DS), so, this means that the 4xVFPv2 part of Digital Foundry is incorrect and actually it only has one VFP?

With that in mind, performance would translate into 3.2 GFLOPS.

I'm not sure if this measurement it's correct, what do you guys think? Also, if someone knows about the VRAM speed of the 3DS or if the New 3DS RAM it's faster it would be sweet!.

 

[Bazooka_Face]

Hi there, I'm curious about some 3DS specs. I think this has been debated enough in the past, but I didn't find any particular numbers on the 3DS CPU actual floating point capabilities.


I know the following about the 3DS specs (New):
ARM11 CPU: 4x MPCore/4x (1?) VFPv2 Co-Processor at 804MHz. Includes 2MB L2 cache (1.25 DMIPS/MHz, but since it has L2 cache let's bump that to 1.4 DMIPS/MHz as a slight boost = 4502 DMIPS)
RAM: 256MB FCRAM (Old 3DS was 3.46GB/s, maybe New 3DS has faster RAM since it's a different chip?)
VRAM: 6MB (I saw at the documents, and actually New 3DS has 6MB of VRAM and not 10MB as Digital Foundry reported, can't find info regarding bandwidth but it's likely higher than main memory)
GPU: PICA200 at 268MHz (6.4 GFLOPS, 20.5 Million polygons per second at 268MHz, calcutated from 15.3 Million figure and 4.8 GFLOPS at 200MHz)

Now, the question would be how much flops per cycle can the CPU do? I'm reading this, and IF I'm not mistaken, the CPU can do 4 instructions per clock, since it's 2 DP MUL MAC, which translanslates into 4 FLOPs.

So... with that, the CPU performance would be: 804MHz (CPU speed) x 4 (Single VFPv2 FLOPs) x 4 (VFPv2 number of Co-Processors) = 12.8 GFLOPS.

However, according to Nintendo, the CPU specs are the following:

As you can see, they don't mention 2x VFPv2, but rather, just a VFP co-processor for the CPU (I assume, the whole processor and not a VFP per core). Also, there's no mention of adding new VFPv2 processors into SNAKE (New 3DS), so, this means that the 4xVFPv2 part of Digital Foundry is incorrect and actually it only has one VFP?

With that in mind, performance would translate into 3.2 GFLOPS.

I'm not sure if this measurement it's correct, what do you guys think? Also, if someone knows about the VRAM speed of the 3DS or if the New 3DS RAM it's faster it would be sweet!.

This wlis all speculation, with no knowledge of the processors low level workings. You cannot use these "mips" to measure/compare the speed or measure the speed of a processor, as the amount of instructions a cpu can do in one second is pretty much useless if the instructions dont get much work done. That's why mips means, meaningless information of processing power. You also cant compare different architectures of CPUs or gpus, as they all are functionally different and perform in different ways.

 

[ghjfdtg]

Nice necro bump. Anyway, if the op will ever read this...

The VFPv2 floating point unit is part of the CPU cores. Each core has its own.

 

[Deleted member 323844]

Core 2 is unused or disabled?

 

[Bazooka_Face]

Core 2 is unused or disabled?

Disabled, and probably removed from the die on later units like the switch did

 

[Deleted member 323844]

Disabled, and probably removed from the die on later units like the switch did

Right. I suspected that.

 

[PabloMK7]

Disabled, and probably removed from the die on later units like the switch did

Core 2 can be used by New 3DS specific games, disabled in other games otherwise. It's not removed in later units lol

 

[Sweater Fish Del]

Is there any homebrew that runs multithreaded to take advantage of two or more cores at once on the New 3DS?

The OP's table says that up to 30% of CPU time on core1 is available for applications. Can this be utilized together with the 100% from core0 or is it either/or? Is there any example homebrew code for this?

Also, is it possible to take over the system cores? A 2DS XL (or 3DS with the 3D slider off, I suppose) has no use for the face-tracking feature, so could core3 be freed up for application use by an exploit?

Or for that matter, could a homebrew bypass the system OS altogether (booting from fastboot3DS) and use all the cores?

 

[PabloMK7]

Is there any homebrew that runs multithreaded to take advantage of two or more cores at once on the New 3DS?

The OP's table says that up to 30% of CPU time on core1 is available for applications. Can this be utilized together with the 100% from core0 or is it either/or? Is there any example homebrew code for this?

Also, is it possible to take over the system cores? A 2DS XL (or 3DS with the 3D slider off, I suppose) has no use for the face-tracking feature, so could core3 be freed up for application use by an exploit?

Or for that matter, could a homebrew bypass the system OS altogether (booting from fastboot3DS) and use all the cores?

There is a bit of multithreading used in this piece of code:
https://github.com/mariohackandglit...05f9f15270cd9102b3/source/updater.c#L237-L282

While one thread is downloading data, the other thread is saving the data to the SD card. You can get some understanding for synchronization tools there. (Keep in mind this isn't using the core1, because it is specified like that in the threadCreate call, but it could be run on any other core just fine.)