Skip to content

armNES 1.4

Choose a tag to compare

@MagnominusGit MagnominusGit released this 20 Nov 23:30
· 0 commits to cb018f3e16f124b73bd874dc0d6f87e12363ecca since this release

armNES 1.4: the last slow boards catch up, picture layers and sound channels can be switched off, and run-ahead and netplay hold on every board.

Updated in place on 20 November 2026: the files below now also run mapper 355 more than twice as fast, without stepping each CPU cycle. If you downloaded 1.4 before, download it again.

What's new

  • Speed. 1.3 left a group of boards slower than the rest: MMC3-style boards with the background at $1000, the FCEUmm boards that stepped a counter every CPU cycle, multicarts reading their ROM through a handler, and a few others. They now run close to a plain cartridge. In a speed sweep of every NES game in my library, about 230 games took four to nine times the median frame time in 1.3. In 1.4 every game is within two and a half times the median.

    Behind it:

    • MMC3-family boards (MMC3, MMC6, RAMBO-1, FK23C, COOLBOY, UNL-8237) with the background at $1000 render whole tiles, and their IRQ is worked out ahead instead of checked every few dots.
    • FCEUmm boards counting CPU cycles run their counters in batches, as the native boards have since 1.3: mappers 40, 42, 43, 50, 65, 67, 83, 105, 106, 142, 178, 222, 264, 303, 308, 330, 359, 368, 406, 417, 446, 451, 540, 548, 555, 561, 562, the FFE boards (6, 17), and FCEUmm's VRC2/4, VRC3, VRC6, VRC7, MMC1, H3001, FME-7 and J.Y. clones.
    • Multicarts whose pages read plain ROM most of the time are mapped straight for the CPU (30, 59, 63, 134, 212, 227, 235, 288, 328, 332, 380, 422, 488, 519, 534, 539, 554, 585).
    • Writes that can't change what the PPU fetches (work RAM, PRG banks, sound registers) no longer catch the PPU up.
    • Mapper 355's PIC microcontroller reads the CPU's address bus itself and runs in batches up to its next port access, worked out from its own program, instead of every CPU cycle. Its games take 3.1 to 3.5 ms a frame on an RG353M (Cortex-A55 at 1.8 GHz), 7.8 before.
    • MMC5 past its IRQ line, MMC5 with a silent pulse, KS-7030 (347), Yuxing (169), mappers 96, 165, 272, 394, 518 and a Dendy's idle lines after the picture.
  • Layer Selection (end of Video, off by default) shows switches for the background, the sprites in front of it and the sprites behind it. Channel Selection (end of Audio, off by default) shows a switch for Pulse 1, Pulse 2, Triangle, Noise, DMC and each voice of the cartridge's sound chip. They change only what's drawn and heard: the game sees every layer and channel, so it plays on the same, and a channel switched off keeps its volume for when it's back on.

  • Netplay with different sound settings. Save states no longer carry the sound output, which follows each player's volume and position settings. Players with different settings saved different states, so RetroArch turned its desync checks off when they joined.

  • Run-ahead goes the same way as without it on every board. Three didn't: 116 (the frame after a load could run on the frame before's hooks), 185 (its disabled CHR read 0 after a load) and 394 (a state from its MMC3 mode loaded over its J.Y. mode kept the wrong handlers; this also went wrong without run-ahead).

  • States across platforms. States of the FCEUmm VRC2/VRC4 boards (14, 116, 222, 351, 446) had a different size on x86-64, 64-bit ARM and 32-bit ARM, so a netplay client on another architecture went its own way after joining.

  • Fixes:

    • Boards that watch the CPU's writes to $2007 (the FFE boards' CHR RAM lock, 195, 252/253, 544) now see them. On 544 a menu's text came out as blank tiles.
    • KS-7030 (347) no longer lets writes into the ROM between its RAM areas.
    • Mapper 355's PIC sees every address the CPU puts on the bus, RAM and registers included, as on the cartridge, not only its reads and writes from $4020.
    • 272 and 394 follow the PPU's timing as the exact model does; 1.3 checked on them every 8 dots and was off.
    • An MMC3 IRQ worked out from partway through a line's sprite fetches could come late. No game in my library showed it.
  • Everything else plays exactly as in 1.3, across all my NES and Disk System games.

Checked: every NES and Disk System game (picture, sound and RAM through 1800 frames with presses) plays as in 1.3 apart from the changes above, in Fast and on a 668-game set in Balanced. Run-ahead went the same way as without it on every game in my library before the last fixes, which don't change what a save state holds, and per mapper with changed accuracy, sound and layer settings. Netplay between an RG353M (64-bit and 32-bit RetroArch) and an x86-64 PC, each side hosting, joined with no CRC warning for games on the boards changed in this release.

Upgrading

  • Save states from 1.3 load, except those of the FCEUmm VRC2/VRC4 boards (14, 116, 222, 351, 446), which may not.
  • Games on 272 and 394 may play on slightly differently from a state, with their timing fix.

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-1.4-*.zip RetroArch: NES, Famicom and Disk System games
armnes_arcade_libretro-1.4-*.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 1.3 for the batched sound chips and the Sunsoft 5B's volume. The README covers options, controls, credits and licenses.