Arcade 60 fps plan

The plan was to run the 68000 arcade games of mame-go (Neo Geo, CPS1, later Sega System 16) at 60 drawn frames a second on the ESP32-S3, with no frameskip. This page says where that stands, what was done and what is left. The day-by-day record (every measurement, the original phases, the logs) is in the repository: docs/archived/arcade-60fps-plan-log.md.

Related pages: Arcade (MAME), 68000 dynarec, JIT plan, Software overview.

Update (2026-10-14): CPS1 resumed

The user put the CPS1 back on the list (priority GBA → CPS1 → Neo Geo). With every frame drawn (bench window 1800 → 4200): Final Fight 34.3 → 37.3 fps, Street Fighter II 22.0 → 32.2, Knights of the Round 24.8 → 31.3. In play, with the frameskip, in 2 minutes of real fighting: Final Fight 60, Street Fighter II 52-53, Knights 52. What did it: the layers drawn in 16-line bands in internal RAM (8-bit and 16-bit games), the arcade flash cache grown to 4.25 MB (Street Fighter II's frames converted on core 1 again), and tiles hidden under fully opaque tiles of later layers left out (adaptive). Details: Optimizations. The table below is the 2026-10-04 state.

Where we were (2026-10-04)

The goal is not reached on any 68000 arcade game, and on this hardware it is not in sight for most of them. What was reached: the games run at or near full speed while about a third of their frames is drawn. Work on the plan was stopped here by the user's decision.

Game	Every frame drawn: core 0 ms a frame (16.7 needed), start → now	In play with the frameskip: game fps / drawn
Metal Slug (Neo Geo)	29.2 → 22.4	58 / 20-21
Metal Slug 2 (Neo Geo, raster)	36.2 → 33.5	46 / 15-16
Final Fight (CPS1)	40.1 → 29.5	55 / 18
Street Fighter II CE (CPS1)	38.9 → 36.1	46 / 15
Golden Axe (System 16)	not measured	54 / 18
Alien Syndrome, Wonder Boy III (System 16)	not measured	58 / 19, 57 / 19-20

• The 68000 came down by a third (Metal Slug 10.8 → 7.0 ms, Final Fight 9.7 → 6.8 in play), not by the half that was the intermediate objective.
• Metal Slug is 6 ms from 60 drawn frames, with 10 ms of video on core 0 and 7 ms of 68000. No measured lead gives that.
• The CPS1 and the raster Neo Geo games are at twice the frame budget.
• The realistic target on this board: the game always at full speed (60 frames of logic a second, clean sound) with 20 to 30 frames drawn. Metal Slug and Final Fight are nearly there, and since 2026-10-05 the System 16 games (54 to 58); Street Fighter II and Metal Slug 2 run at 45 to 50.

Launch times: the same arcade game again starts in 9 to 13 s; after another arcade game it takes 20 to 28 s, because the flash cache (4 MB, 4.25 MB since 2026-10-14) holds one game and is rewritten (the screen says "game cache NN%").

What was done

In the order it was done; each step was proven frame by frame on the PC harness, then measured on the board in play.

Step	Scope	What it gave
Palette fast paths, hot code in IRAM, aligned 16-bit accesses, inline RAM/ROM path in Musashi	all MAME games	about 2 ms a frame together
The display task reads MAME's 8-bit bitmap itself; change test per line; branch-free scalers	every emulator	display task 12.2 → about 4.5 ms of CPU
Neo Geo frame drawn in 16-line bands in internal RAM, handed to the display as they finish	Neo Geo	video off the PSRAM frame bitmap
Sound board (Z80 + FM) on core 1; chips rendered at 16 kHz, doubled to the 32 kHz output	Neo Geo, CPS1	1.5 to 2 ms a frame, not audible
Exact skips of the 68000's waiting code (counting loops, idle turns)	Neo Geo, CPS1, System 16	68000 −3 to −4.7 ms where they fire
Card pages through a DMA-capable buffer	Neo Geo sprite and sample pagers	17 → 7 ms and 12 → 5 ms a read
CPS1: clears cut to the visible area; 8-bit frame handed to the display task	CPS1	Final Fight 40.1 → 29.5 ms; 48.6 → 55.3 fps in play
SD reads into PSRAM several KB per command instead of one sector	every emulator	SNES 4 MB launch 10.9 → 5.8 s, GBA 8 MB 10.8 → 6.8 s
Resume at boot fixed; CPS1 states smaller; SF2 states borrow a video cache	Neo Geo, CPS1	saves work on every tested game
Screenshot from a frame-boundary copy	mame-go (not the Neo Geo band mode)	right pictures on the indexed path

Also added on the way: a loading percentage in every emulator, a CPS-1 tab and covers in the launcher, Sega System 16 in mame-go, OutRun through Cannonball, Arcade 3D Racing.

Tried and dropped (measured; do not retry without a new reason):

• the 68000 dynarec: exact, slower than the interpreter here (see 68000 dynarec);
• PSRAM at 120 MHz and the LCD bus at 25 MHz: out of spec, excluded by the user;
• 16-bit bitmaps for the Neo Geo, sprites on core 1, a faster sprite plotter, the program's first megabyte in PSRAM, FAME/C, lazy cycle counting;
• FatFs with 512-byte sectors for internal RAM: no gain;
• a cache of decoded CPS1 tiles on the card: slower than the zip;
• an even frameskip cadence (one frame in N): the same drawn frames, off (file skip_even);
• a screenshot in the Neo Geo band mode by running one more frame: it hung the game.

What is left

In the order of what it could give.

1. System 16: done on 2026-10-05 (sound board on core 1, frames handed to the display, idle skips stepping aside where the driver has its own idle hack, full screen): Golden Axe 50 -> 54 fps in play, Alien Syndrome 50 -> 58, Wonder Boy III 49 -> 57. Left: Shinobi and Altered Beast (no set at hand); the System 16 save states, which keep only the CPUs and RAM.
2. The CPS1's renderer on core 1: tried on 2026-10-13 and dropped (Final Fight 34.7 -> 30.0 fps; core 1 shares the PSRAM and already runs the sound board and the display). The bands in internal RAM were kept instead.
3. Street Fighter II: its PSRAM shortage was solved on 2026-10-14 by the 4.25 MB flash cache (frames converted on core 1 again, play 47 -> 53); the 16-bit games (Knights of the Round) are now in bands too (play 49 -> 52). Left for both: Street Fighter II's row-scrolled floor copy and the sprites, walked in full for every band.
4. The Neo Geo raster games' 16-bit video path (Metal Slug 2, KOF '95): 16 ms.
5. A screenshot for the Neo Geo band mode, with a PC reproduction first.
6. Open defects: a CPS1 state is not an exact continuation; the dead buttons of 2026-10-03 were never explained (the binary and the card were cleared). robby.zip resuming black was fixed on 2026-10-05 (the Astrocade video registers are now in the save state), and Street Fighter II's saves complete again.

What the hardware would change: a module with 32 MB of flash and 16 MB of PSRAM removes the flash cache rewrite, the lack of room for apps (64 KB left) and Street Fighter II's memory limits, not the frame rates. See Plan D.

How to work on it

• Measure in play. MAMEBENCH=2 (Neo Geo) and MAMEBENCH=4 (CPS1, System 16) insert a coin, press START and play a script; MAMEPROF=1 samples core 0, NEOPROF=1 times the parts. The attract loop and the intro screens flatter every figure. Details: scripts/mamebench/README.md.
• Prove a change on the PC first. scripts/neogeo_frames.py compares pictures and sound frame by frame (ref, compare, turn, count).
• Compile on the Mac, test on the board: scripts/retro_go_build.sh <app> (for mame-go with the play options, or the band code is not compiled).
• Switch files in /sd/retro-go/mame/ turn single changes off for a comparison: cps1_noturn, cps1_fullclear, cps1_noindexed, cps1_indexed, sys16_noturn, skip_even, neo_nocount.
• After any bench or profile run the play build goes back (scripts/mamebench/board_install.sh, which refuses a bench binary), the switch files are removed, and coin + START are checked on Metal Slug: a bench build ignores the gamepad.
• Limits: 64 KB of flash left outside the app partitions, about 18 KB in the launcher's, 0 to 5 KB of internal RAM in Metal Slug.