A new Nintendo 3DS emulator core for Apple hardware.
Our goal is the smoothest 3DS experience on Apple hardware without CPU JIT. Cara brings a performance-first approach to emulation on iPhone and iPad: custom GPU rendering, efficient memory use, and gameplay without CPU JIT. Open source, with a native shell of its own and a reusable core designed for Delta integration.
In a recorded Xenoblade Chronicles 3D opening replay on an M4 iPad, Cara reached 99.5% emulation speed using 758 MiB peak sampled memory—about 40–45% less memory than the two Azahar-based apps in the same comparison.
| Build | Emulation speed | Peak sampled memory |
|---|---|---|
| Cara GLES3 | 99.5% | 758 MiB |
| Manic / Azahar | 97.4% | 1,373 MiB |
| Azahar / MoltenVK | 84.6% | 1,261 MiB |
All three completed the opening replay at native resolution, with sound enabled, CPU JIT disabled, and 100% guest CPU clock. This is a whole-app comparison of one scene; performance and compatibility vary by game. See the full results, build revision, and methodology.
| Focus | The Cara approach |
|---|---|
| Built for Apple hardware | Custom GPU rendering and native presentation for iPhone and iPad. |
| More work on the GPU | GPU vertex processing and packed-data decoding reduce work on the CPU. |
| Less repeated work | Prepared-input reuse and fewer CPU/GPU transfers target the costs that slow gameplay down. |
| Memory-conscious design | Bounded graphics caches and data reuse keep memory efficiency central to performance tuning. |
| Play without CPU JIT | Cara extends Azahar FastInterp with static arithmetic fast paths, decoded-block caching and execution-budget fixes inside the signed app. |
| A core and a shell | A reusable core developed for Delta integration, alongside Cara’s standalone iOS app. |
Citra MMJ, weihuoya’s performance-focused Android fork, gave Cara a mobile rendering foundation to build on. We carry that work forward in three key areas:
- Mobile GPU rendering. MMJ’s OpenGL ES rendering path is the starting point for Cara’s Apple-focused graphics engine. Cara adds custom GPU vertex processing, packed-data decoding, and native screen presentation.
- Shader generation and reuse. MMJ generates and reuses GPU shaders and programs to avoid repeated compilation. Cara adapts that shader foundation for Apple GPUs and extends it with its own rendering enhancements.
- Graphics data reuse. MMJ’s texture and render-target caching provides the foundation for reusing graphics resources. Cara takes that focus further with prepared-input reuse, bounded graphics caches, and fewer CPU/GPU transfers.
Azahar FastInterp supplies the interpreter foundation. Cara extends it with static arithmetic fast paths, faster instruction-timing lookup, and an eight-way decoded-block cache on iOS. Cara also fixes decode-record reuse and execution budgets across chained blocks. Guest CPU instructions remain interpreted inside the signed app, without runtime CPU code generation.
Cara combines these execution changes with its Apple graphics pipeline, bounded memory and upload caches, a reusable core, and its own native shell.
Cara’s progress has come from understanding more of the machine: how the guest CPU produces graphics commands, who owns each byte of shared memory, when work becomes visible, and how the host presents it. Changing the graphics API alone would leave that machinery to solve. Metal remains a future direction; the current work makes the engine underneath it better.
The GLES3 results above demonstrate what this approach can deliver. They compare whole applications, so they do not isolate the performance of one graphics API. The stronger evidence comes from changing one mechanism at a time:
| Finding | What changed or what we learned |
|---|---|
| Keep graphics data on the GPU | GPU alias remapping lets compatible surfaces share GPU-resident data while preserving guest-visible bytes. A matched Luigi laboratory replay fell from roughly 24.5 to 16.7 ms per refresh, eliminating six downloads per refresh with matching final pixels. Results |
| Preserve memory ownership | Targeted cache cleanup and alias invalidation preserve coherence without flushing unrelated dirty surfaces. Partial writes, overlapping views and CPU reads need explicit correctness checks. Residency research |
| Fix host pacing | Bounded pacing recovered near-full speed in a matched Xenoblade replay while preserving guest event timing, draw counts and final pixels. Results |
| Follow the guest’s work | Stories traces distinguish command submission, draw work, completion events and display refresh. Offline instruction inspection identified copying, visibility tests and a tick-poll pacing loop; their execution counts alone do not establish CPU cost or justify skipping them. Trace findings |
| Measure the waiting work | A bounded Stories probe measured about 0.69 ms of instrumented CPU time per sampled refresh in the fine polling episode, including conversions and its tick query. It does not explain the whole slowdown; observer overhead remains in the figure and outside-interpreter scheduler cost is still unmeasured. Findings |
| Keep experiments accountable | Staging only shader-used inputs removed 14% of configured input registers in a Stories comparison, but interleaved data kept payload bytes unchanged and there was no measured speed gain. The option stays off. Results |
These research results describe the development builds recorded in each report. The target is full speed at 100% guest CPU with sound and no CPU JIT. Some Stories scenes still fall short. Underclocking can help playability, but it also changes how much work the guest produces; it cannot establish a full-clock gain. Hardware diagrams and register research guide the investigation, while matched checkpoints, memory-coherence fixtures and pixel comparisons test each change.
Explore the architecture, hardware evidence, and research references.
- Your games, through Files. Open extracted game files from the file browser.
- Your controls. Touch controls and customizable Delta-compatible skins.
- Your progress. Per-game save-state slots and in-game save syncing through a selected iCloud Drive folder.
- Your experience. GPU-accelerated graphics, native stereo audio, and an optional adaptive guest CPU clock for demanding scenes.
The standalone Cara app is the playable build provided here. Delta core integration is a development target.