The macstudio_m5ultra_96-ud-iq-xs configuration delivers a balanced performance profile on the Mac Studio M5 Ultra, characterized by high-speed text ingestion and moderate generation rates. This setup leverages the device's substantial 96 GB unified memory to accommodate the model while maintaining a manageable power footprint of 360 watts during active inference tasks.
Measuring Throughput and Memory Utilization
To verify the reading throughput of 2284.8 tokens per second, you can monitor the input processing speed using the Strata client logs. This metric indicates how quickly the engine can ingest context, which is critical for long-context applications. The system utilizes the full 96 GB of VRAM, ensuring that the quantization does not cause out-of-memory errors even with large context windows. You can confirm this allocation by observing the memory usage metrics in the system monitor while the Strata engine is running.
Analyzing Generation Speed and Efficiency
The writing throughput of 64.8 tokens per second represents the actual generation speed for text output. While this is not the fastest possible generation rate, it is sufficient for most interactive applications and research tasks. The power consumption of 360 watts reflects the energy required to sustain this level of computation. For users concerned about operational costs, this draw is consistent with high-performance local AI workloads on Apple Silicon architecture.
Cost and Performance Summary
The fit is rated as Good, indicating that the quantization level strikes an effective balance between model fidelity and resource consumption. This configuration is ideal for users who prioritize context handling over raw generation speed.
| Metric | Value | Unit |
|---|---|---|
| VRAM Capacity | 96 | GB |
| Reading Throughput | 2284.8 | Tokens/Second |
| Writing Throughput | 64.8 | Tokens/Second |
| Power Consumption | 360 | Watts |
ud-iq-xs is a quantization format that reduces model precision to optimize memory usage and inference speed without significantly compromising output quality.