# RTX 5090 vs M5 Ultra: Which AI Powerhouse Wins?

## Executive summary

This video compares the Apple M5 Ultra Mac Studio and the NVIDIA RTX 5090 for running large language models (LLMs) locally. The M5 Ultra offers significantly higher unified memory capacity (up to 256 GB) and memory bandwidth (1.2 TB/s), while the RTX 5090 boasts superior raw compute power (419 TFLOPS) and higher memory bandwidth (1.8 TB/s). Performance benchmarks for Qwen3.8-27B show the RTX 5090 leading in raw token generation speed, especially with Multi-Token Prediction (MTP) enabled, but the M5 Ultra's massive RAM capacity makes it superior for running larger or unquantized models.

## Key takeaways

- Memory Capacity Advantage: The M5 Ultra can be ordered with up to 256 GB of RAM, allowing users to run much larger models compared to the RTX 5090's 32 GB of VRAM.
- Raw Compute Power: The RTX 5090 achieves approximately 419 TFLOPS, significantly higher than the M5 Ultra's 110 TFLOPS.
- Token Generation Speed (MTP On): With Multi-Token Prediction (MTP) enabled, the RTX 5090 achieves 136 tokens per second, compared to the M5 Ultra's 100 tokens per second.

## Technical details

- Memory Bandwidth: M5 Ultra: 1.2 TB/s; RTX 5090: 1.8 TB/s.
- Compute Performance (TFLOPS): M5 Ultra: ~110 TFLOPS; RTX 5090: ~419 TFLOPS.
- Power Consumption (Sustained Peak Load): M5 Ultra: ~214 watts; RTX 5090: 300 to 575 watts.
- Qwen3.8-27B Benchmarks (Prompt Processing): RTX 5090: ~3,000 tokens/second; M5 Ultra: ~1.7K tokens/second.

## Practical implications

- For running models locally, the RTX 5090 offers superior raw inference speed (tokens/second) due to higher bandwidth and TFLOPS, provided the model fits within 32 GB.
- The M5 Ultra's large unified memory (up to 256 GB) is critical for running larger, unquantized, or Mixture of Experts (MoE) models that exceed the 32 GB limit of the RTX 5090.
- Inference speed (tokens/second) is fundamentally determined by bandwidth divided by model size.

## Topics

Local LLM Inference, Hardware Benchmarking, AI Compute Architecture, M5 Ultra, RTX 5090, Qwen3.8-27B, RTX 5090 llama.cpp runs of Qwen3.8-27B Q4_K_M, MacStories M5 Ultra review

Source: https://www.youtube.com/watch?v=qaP4pymNrEc