Idea: # Significant Performance Difference Between Longshot and NX2navi #1912

Open
opened 2026-11-20 20:29:51 +00:00 by bigwiggler · 1 comment

The idea

Significant Performance Difference Between Longshot and NX2navi

Longshot is delivering approximately 50% lower performance than NX2navi in the same Nintendo Switch 2 game, and Longshot also exhibits noticeable stuttering that is not present in NX2navi.

This is not a minor difference. Both projects are implementing Switch 2 emulation through Vulkan, and both have substantial replay optimizations. NX2navi demonstrates that this workload can run significantly faster and with smoother frame pacing.

What I am experiencing

  • NX2navi: Approximately 50% faster performance, with consistent frame pacing and no noticeable stuttering.
  • Longshot: Significantly lower FPS and recurring stuttering in the same game and environment.

Why I believe this deserves investigation

Looking through the two projects' Vulkan implementations, I found that both already implement several similar optimizations:

  • Parallel command-buffer replay.
  • Batched Vulkan queue submissions.
  • Background Vulkan recording.
  • GPU memory mirroring and synchronization mechanisms.

Longshot already has parallel replay and submission batching, so simply adding those features is not the answer. The important question is why its implementation is performing so much worse.

Some areas of the Longshot implementation that deserve profiling include:

  1. Replay critical path: Determine whether command processing, replay coordination, or synchronization is unnecessarily blocking the game thread.
  2. GPU memory mirroring: Investigate whether timeline waits and dirty-page synchronization are causing excessive GPU stalls. Longshot's own source comments document a configuration where mirror-related waits substantially reduced performance.
  3. Queue submission overhead: Compare Longshot's batching and recorder implementation against NX2navi's approach, including the cost of synchronization between threads.
  4. Frame pacing: Identify what is causing the recurring stutters when NX2navi handles the same workload smoothly.

What I am requesting

Please investigate the performance difference between Longshot and NX2navi, identify the bottleneck, and implement the necessary optimization or fix.

I am not asking for a collection of configuration tweaks to work around the problem. I am reporting a substantial performance difference between two implementations targeting the same hardware and workload.

A roughly 50% performance gap, combined with stuttering that does not occur in NX2navi, suggests there is meaningful optimization work to be done in Longshot's Vulkan execution or synchronization path.

Anything else

No response

### The idea # Significant Performance Difference Between Longshot and NX2navi Longshot is delivering approximately **50% lower performance than NX2navi** in the same Nintendo Switch 2 game, and Longshot also exhibits noticeable stuttering that is not present in NX2navi. This is not a minor difference. Both projects are implementing Switch 2 emulation through Vulkan, and both have substantial replay optimizations. NX2navi demonstrates that this workload can run significantly faster and with smoother frame pacing. ## What I am experiencing - **NX2navi:** Approximately 50% faster performance, with consistent frame pacing and no noticeable stuttering. - **Longshot:** Significantly lower FPS and recurring stuttering in the same game and environment. ## Why I believe this deserves investigation Looking through the two projects' Vulkan implementations, I found that both already implement several similar optimizations: - Parallel command-buffer replay. - Batched Vulkan queue submissions. - Background Vulkan recording. - GPU memory mirroring and synchronization mechanisms. Longshot already has parallel replay and submission batching, so simply adding those features is not the answer. The important question is why its implementation is performing so much worse. Some areas of the Longshot implementation that deserve profiling include: 1. **Replay critical path:** Determine whether command processing, replay coordination, or synchronization is unnecessarily blocking the game thread. 2. **GPU memory mirroring:** Investigate whether timeline waits and dirty-page synchronization are causing excessive GPU stalls. Longshot's own source comments document a configuration where mirror-related waits substantially reduced performance. 3. **Queue submission overhead:** Compare Longshot's batching and recorder implementation against NX2navi's approach, including the cost of synchronization between threads. 4. **Frame pacing:** Identify what is causing the recurring stutters when NX2navi handles the same workload smoothly. ## What I am requesting Please investigate the performance difference between Longshot and NX2navi, identify the bottleneck, and implement the necessary optimization or fix. I am not asking for a collection of configuration tweaks to work around the problem. I am reporting a substantial performance difference between two implementations targeting the same hardware and workload. **A roughly 50% performance gap, combined with stuttering that does not occur in NX2navi, suggests there is meaningful optimization work to be done in Longshot's Vulkan execution or synchronization path.** ### Anything else _No response_
Member

Possibly related, for context:

  • issue #1361 Game: Higher FPS than NX2Navi feels less smooth — Same NX2Navi comparison and perceived smoothness, but this report alleges a much larger FPS deficit.
  • issue #402 Deck: the frame is CPU-bound in the submission path, not GPU-bound (16 fps at 50% internal, GPU-bound wait 0.4 ms) — Profiling documents CPU and synchronization/replay bottlenecks, but on Steam Deck rather than the new comparison.
  • issue #1639 Game: Inconsistent performance and severe stuttering in Hyrule Castle Town / Temple of Time — Broader reports of low FPS and stuttering across areas; cause not established.
  • PR #1866 perf(vk): the recorder takes a whole part and submits it, so the render thread never waits at a part's end — Merged Vulkan recorder change addresses replay hand-off overhead; not a demonstrated fix for this claimed overall gap.
Possibly related, for context: - issue #1361 Game: Higher FPS than NX2Navi feels less smooth — Same NX2Navi comparison and perceived smoothness, but this report alleges a much larger FPS deficit. - issue #402 Deck: the frame is CPU-bound in the submission path, not GPU-bound (16 fps at 50% internal, GPU-bound wait 0.4 ms) — Profiling documents CPU and synchronization/replay bottlenecks, but on Steam Deck rather than the new comparison. - issue #1639 Game: Inconsistent performance and severe stuttering in Hyrule Castle Town / Temple of Time — Broader reports of low FPS and stuttering across areas; cause not established. - PR #1866 perf(vk): the recorder takes a whole part and submits it, so the render thread never waits at a part's end — Merged Vulkan recorder change addresses replay hand-off overhead; not a demonstrated fix for this claimed overall gap.
Sign in to join this conversation.
No milestone
No project
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
MagicBeanDev/Longshot#1912
No description provided.