Define the pass condition
Answer the following question while evaluating Neural Frame Rate Upscaling (NFRU):
Can NFRU meet the presentation target of the game within its quality, latency, power, and thermal limits?
Use the Project Moku case study to learn how to test a game. Then, repeat the steps from the case study with your game.
Set the thresholds before looking at the result. The evaluation guidance in Arm Neural Technology Playbook - Evaluate explains the wider tradeoffs.
| Area | What to record | What it defines |
|---|---|---|
| Presentation | Render FPS, present FPS, frame-time distribution | Target cadence and pacing variance |
| Cost | GPU and neural time, power, thermals | Maximum cost and sustained duration |
| Quality | Captures from difficult content | Acceptable artifacts in motion |
| Responsiveness and reach | Latency and device coverage | Latency limit and fallback policy |
Run one controlled pass
Complete one controlled pass on your game:
- Capture a baseline: Record the same short action with NFRU disabled and enabled under fixed conditions.
- Stress the content: Use the Project Moku occlusion test , Project Moku particle-effects test , and Project Moku lighting-change test to select relevant scenes from your game.
- Measure execution: Follow the Streamline and RenderDoc workflow to confirm neural activity and inspect unexpected output.
- Check presentation: Compare render FPS with present FPS and use the frame-pacing analysis to choose a sustainable target.
Keep the model unchanged. First determine whether a failure comes from inputs, content, runtime cost, or pacing.
Profile graph operation on NX
On a supported neural accelerator (NX) target, Streamline 9.9 connects a slow frame to the graph operation behind it. Review Profile Arm Neural Technology workloads with Streamline to complete the following profiling workflow:
- Prepare your device and debuggable application.
- Collect a capture through the GUI.
- Inspect neural accelerator activity.
Select the interval in the Mali Timeline view. Then, use the Dispatch Graph Workloads view to find expensive operations, outlier dispatches, and shader fallbacks. Repeat the same capture after each change. To learn more about the graph-analysis views and CSV export options, see the Streamline neural-graphics profiling overview .
Make a decision
You have enough information now to make one of the following decisions:
- Continue if the integration meets every threshold.
- Iterate if the failure is reproducible and has a credible fix.
- Stop if quality, latency, sustained cost, or fallback misses the product target.
What you’ve accomplished and what’s next
You now have a controlled NFRU result and a clear decision.
Next, you’ll consider whether the more demanding Neural Super Sampling and Denoising (NSSD) route belongs in your plan.