Skip to content
Michele Primavera edited this page Nov 18, 2026 · 2 revisions

Hardware notes

While building Maelle, we found many things about the DS, the DSi, their system software and the Wii U that weren't written down anywhere, or were written down wrong. These pages collect them for anyone working on an emulator, homebrew, or the Wii U.

Page What's in it
ARM CPUs on the DS and DSi Instruction edge cases measured on real ARM9 and ARM7 cores, CPU speeds, caches, and the timing races games depend on
DS hardware 2D, the 3D engine, the display FIFO, the geometry FIFO, slot-2 accessories, cartridge saves, the GBA cart RTC
DSi hardware NDMA, the microphone, MBK9, SCFG, SD/MMC, AES, the MCU, the touchscreen controller, cameras, GPIO, DSP, DSi cartridges
DSi system software What the DSi menu and launcher set up for a title, and what DSi Sound, DSi Camera, Flipnote Studio and System Settings expect
Games that depend on hardware details Games that break if one specific detail is wrong
Wii U homebrew notes Espresso, Cafe OS threads, Aroma, GX2 and Latte shaders, sockets, the SD card, audio, the GamePad

How sure each note is

Each note is marked with where it comes from:

  • Measured: a test ROM was run on a real DS, a DSi, or both, and saved its results. The ROMs are in tests/hardware, and their README explains how to run them and read the results.
  • Software: real system software or a retail game needs it to work, and it was checked on a Wii U. This shows the behavior is needed, but nobody measured the hardware directly.
  • Unverified: taken from other emulators or worked out from code, and not checked against hardware.

Unless a note says otherwise, "a DSi" means a DSi running a title in DSi mode. Commit links point to the change in Maelle where each one was made.

Where other sources were wrong

These are the places where a common source, or NooDS before this fork, did something different from real hardware.

Topic What the source had What the hardware does How we know
NDMA reserved destination step (3) melonDS: increment Same as fixed Measured on a DSi
MBK9 top byte Cart headers fill it with WRAMCNT, and we first thought it locked shared WRAM blocks Reads back 0 and does nothing; every block follows WRAMCNT Measured on a DSi
ARM7 UMULL/UMLAL early finish NooDS: also on top bytes of all ones, like the signed multiplies Only on top bytes of zeros Measured on a DSi (all 24 cases)
STM with the PC in the list NooDS: instruction address + 8 Instruction address + 12, on both CPUs Measured on a DS
STRD with writeback, base in the pair NooDS: stores the new base Stores the old base Measured on a DS (ARM9)
LDM^ loading the PC NooDS: takes THUMB from bit 0 of the loaded PC Takes THUMB only from the SPSR; with no SPSR, stays in the current state Measured on a DS
MSR in User mode NooDS: could write the I, F and T bits Writes only the flags Measured on a DS
MSR with an invalid mode NooDS: kept the value as written Forces mode bit 4 (00 becomes 10, 07 becomes 17, and so on), for the CPSR and the SPSR Measured on a DS
Reserved CPSR/SPSR bits NooDS: kept them Not stored; the ARM7 also has no Q flag Measured on a DS
MRC with a missing coprocessor NooDS: ignored, or used CP15 Undefined instruction exception (any MRC/MCR on the ARM7, coprocessor 0 on the ARM9) Measured on a DSi
MRC p15 into the PC NooDS: wrote the PC Sets N, Z, C, V from the value's top four bits and leaves the PC alone, as ARMv5 says Measured on a DSi
LDRT/STRT NooDS: did nothing Load or store and write back, like post-indexed LDR/STR Measured on a DS
THUMB NEG NooDS (since 2023): always cleared V Sets V for NEG 0x80000000 ARM architecture
RSCS NooDS: C and V could never go back to 0 Normal flags ARM architecture
ARM9 byte writes to palette RAM NooDS: wrote the byte Ignored Measured on a DS
ARM7 BIOS speed NooDS: half speed, like the rest of the ARM7 No wait states, so full speed Software (Flipnote Studio); DSi speed test
ARM9 code fetches NooDS: free Line fills cost about 22 ticks of 33 MHz on a miss Measured on a DSi
CLEAR_DEPTH at reset NooDS: 0 Games that never write it need the farthest depth (0x7FFF), as melonDS uses Software (Plants vs. Zombies)
BG2/BG3 affine parameters at reset NooDS: all 0 Identity matrix (PA = PD = 0x100) Software (bitmap-mode homebrew)
DIVCNT division-by-zero bit Maelle (briefly): computed only when the result was read Must be correct when DIVCNT is read too Software
DSP nearest-neighbour scaler melonDS: steps down by the rectangle's width The picture's width (only matters for rectangles narrower than the picture) Unverified: from reading the code
DSP bilinear scaler melonDS: loads the same line twice Blends with the line below Unverified: from reading the code
SD/MMC writes NooDS: fread into the data, so DSi NAND and SD writes were lost Written Software

Clone this wiki locally