Skip to content

armNES 1.3

Choose a tag to compare

@MagnominusGit MagnominusGit released this 19 Nov 23:47
· 16 commits to main since this release

armNES 1.3: every board that stepped each CPU cycle now runs in batches, two to four times faster, and the Sunsoft 5B can be heard.

Updated in place on 20 November 2026: the files below now also batch the IRQ-only boards, the karaoke board and the NSF player, fix the J.Y. boards' IRQ timing and the 5B's volume. If you downloaded 1.3 before, download it again.

What's new

  • Every board that stepped its sound or IRQ counter every CPU cycle now runs in batches, as the Disk System has since 1.2. Stepping each cycle kept the whole core on its slow path, so these games cost two to four times a plain cartridge's frame. On an RG353M (Cortex-A55 at 1.8 GHz), ms a frame:

    Board Mapper 1.2.1 1.3
    VRC7 85 10.3 3.4
    VRC6 24, 26 9.7 2.1
    Sunsoft FME-7 with the 5B 69 9.9 2.5
    Namco 163 19 10.9 2.4
    MMC5 5 11.1 6.2
    VRC4 21, 23, 25 4.2 2.0
    VRC3 73 4.3 2.0
    Jaleco SS88006 18 4.4 2.2
    Bandai FCG 16, 153, 157, 159 4.3 1.6
    J.Y. Company 90, 209, 211 5.0 2.9
    Tengen RAMBO-1 64, 158 5.0 2.7
    YM2413 karaoke 515 8.9 2.1
    NSF player, a file that waits for its play timer 14.4 0.9

    Mapper 127 and Konami's QTa adapter (547) are batched too. That leaves room on 1.5 GHz Cortex-A53 handhelds, where the busiest of these games could drop below 60 fps.

  • A program waiting for an interrupt in a jump to itself has those cycles skipped in one go, as polling loops on $2002 already were. NSF files gain most from it.

  • Sunsoft 5B: its three channels were about 8 dB too quiet and hard to hear next to the 2A03. The 5B's DAC is now its 5-bit, 1.5 dB-a-step curve (NESdev), and a channel at full volume is about 8 dB louder than a 2A03 square, as rainwarrior measured on a real 5B (Nestopia mixes it the same). The amplifier's compression at the loudest volumes isn't modelled; music at the volumes games use doesn't reach it.

  • J.Y. Company boards (90, 209, 211, 35) counting the PPU's A12 line raise their IRQ on time. It came late before, so some screens split a few lines off: among those games, 18 now match the exact timing model through the first 1800 frames and the rest run closer to it. RAMBO-1 (64) games run closer to it too.

  • A save state made before the first frame no longer loses a few CPU cycles when loaded. With run-ahead, a game could drift a few cycles from where it would have been without it.

  • Everything else plays exactly as in 1.2.1, across all my NES and Disk System games.

Netplay checked on an RG353M (64-bit and 32-bit RetroArch) with an x86-64 PC, each side hosting, with games for each of these boards and the Disk System: no state differences at any check.

Upgrading

  • Save states from 1.2.1, 1.2 and 1.1 load. Games on J.Y. boards using A12, and on RAMBO-1, may play on slightly differently from a state, with the timing fix.
  • Sunsoft 5B audio sounds louder: the chip is mixed at the level it has.

Known gaps

  • The 41 boards that come from FCEUmm with a per-cycle counter still step each cycle. Their counters live inside FCEUmm's code; they're mostly pirate and multicart boards.
  • Games that draw the background from the $1000 pattern table on an MMC3-style board, and NSF files whose code runs from $F000-$FFFF, stay slower than most.

Downloads

File For
armNES-x86_64.AppImage, armNES-aarch64.AppImage The standalone app on Linux
armNES-arm64.apk The standalone app on Android
armnes_libretro-1.3-*.zip RetroArch: NES, Famicom and Disk System games
armnes_arcade_libretro-1.3-*.zip RetroArch arcade folders: MAME's VS. System and PlayChoice-10 sets

The core zips come for x86-64, 64-bit ARM (aarch64) and 32-bit ARM (armv7).

See armNES 1.2 for the cheat cartridges and the expansion sound fixes. The README covers options, controls, credits and licenses.