Customize only for a measured reason
Start with the published model. Customize it only when the runtime exposes a reproducible quality or performance gap. To know when to customize a model, see the model workflow in Arm Neural Technology Playbook - Evaluate .
Design with NX in mind
Understand how the GPU Execution Engine and neural accelerator (NX) perform the following tasks:
- Divide a workload
- Choose precision and quantization deliberately
- Limit avoidable data movement and execution transitions
- Connect model design to measurement, conversion, and deployment
For more information, see Neural Network Model Development Best Practices for GPUs with Neural Accelerator . Use the Arm GPU Best Practices Developer Guide alongside it to optimize the graphics workload that runs with your neural model.
Use Model Gym for NSS and NFRU model workflows
Neural Graphics Model Gym supports Neural Super Sampling (NSS) and Neural Frame Rate Upscaling (NFRU) model workflows. The Model Gym covers the following workflows:
- Training
- Fine-tuning
- Evaluation
- Quantization
- Export for ML extensions for Vulkan
Use Model Gym in four steps:
- Reproduce the issue with a prebuilt model and fixed data.
- Prepare a fixed training and evaluation dataset that represents the measured issue.
- Follow Train and evaluate NFRU models using Model Gym to train or fine-tune an NFRU model. Then, compare quality before and after quantization.
- Follow Prepare models for Arm Neural Technology to export and validate the artifact.
Connect model estimates to measured results
Compare each candidate against the same data. Record quality before and after quantization, and confirm the exported model runs. For a supported model input, use MLIA compatibility checks and NX performance estimates to investigate model cost before deployment. Keep the target profile and configuration fixed when comparing candidates.
When you have supported NX hardware, follow Profile Arm Neural Technology workloads with Streamline to measure the integrated workload. An estimator result doesn’t establish image quality, frame pacing, power, or thermal behavior in your game.
Change one variable at a time, and validate each hardware-aware design choice against your workload.
What you’ve accomplished and what’s next
You now have a bounded model workflow, from a reproducible gap through training, quantization, and export.
Next, you’ll use the model toolkit to connect compatibility checks and performance estimates to device evidence.