On-Device Multimodal Runtime Performance Matrix
Evaluate on-device vision-language model execution trade-offs across mobile chipsets, thermal envelopes, and quantization tiers.
Use this template when architecting local multimodal generative AI features in native Android apps. It produces a detailed matrix comparing inference latency, memory pressure, and visual prompt fidelity across device tiers.
Role: Principal Android Machine Learning Architect specializing in on-device neural acceleration and generative vision pipelines.
Context
- Target Chipset Architectures: {{target_soc_families}}
- Memory Allocation Ceiling: {{max_memory_footprint}}
- Vision-Language Model Selection: {{multimodal_vision_model}}
- Quantization Specification: {{quantization_profile}}
- Device Thermal Budget: {{thermal_budget_celsius}}
- Prompt Token Context Budget: {{prompt_context_length}}
Task
Generate a comprehensive technical evaluation matrix that maps on-device multimodal runtime performance across hardware profiles, providing concrete integration guidelines for local image understanding and generation.
Method
- Analyze the throughput limits of {{multimodal_vision_model}} running under {{quantization_profile}} across each family in {{target_soc_families}}.
- Calculate memory bandwidth saturation and peak RAM usage against {{max_memory_footprint}} during full visual context ingestion.
- Measure thermal throttle onset curves when sustaining generation loops within {{thermal_budget_celsius}}.
- Map prompt token ingestion rates at {{prompt_context_length}} against NNAPI and vendor-specific NPU execution delegates.
- Benchmark fallback routes to GPU and CPU when NPU delegate initialization fails or encounters unsupported ops.
- Evaluate image conditioning artifacts caused by aggressive low-bit integer quantization.
- Synthesize findings into structured evaluation matrices covering latency, memory, thermal impact, and visual fidelity.
Constraints
- MUST evaluate specific vendor delegates (Qualcomm QNN, MediaTek NeuroPilot, Google Tensor).
- MUST NOT recommend execution paths that breach {{max_memory_footprint}} on minimum baseline devices.
- Benchmark metrics MUST be expressed in concrete units (ms per token, MB allocated, °C rise per minute).
- Keep implementation guidance strictly compatible with modern Android NDK and LiteRT/MediaPipe standards.
Output format
Provide a technical specification structured as follows:
- Section 1: Hardware Execution Matrix (Markdown table comparing SOCs, NPU delegates, TTFT, and continuous generation latency)
- Section 2: Memory & Thermal Profile Matrix (Markdown table detailing baseline RAM, peak allocation, and throttling thresholds)
- Section 3: Quantization vs. Visual Fidelity Trade-Off Matrix (Markdown table mapping accuracy loss against quantization levels)
- Section 4: Architectural Recommendation Summary (300 words maximum)
Self-review
- Confirm all 6 context variables are referenced and properly scoped.
- Verify that every matrix table contains concrete Android-specific profiling parameters.
- Ensure no generic cloud fallbacks supersede the on-device constraint validation.
Explicit role, a named task, and discrete steps the model can follow.
Background, inputs and variables the model needs before it starts.
Hard boundaries — what the model must and must not do.
A named, field-level shape for the response.
Ordered work items that force analysis before an answer.
Length and structure that travel across frontier models.
Signal density — instruction weight without padding.
Documented variables so the scaffold adapts to new inputs.
Quality bar, assumptions and behaviour when inputs are thin.
How much real usage the template has behind it.