RTX 5090 Laptop GPU Beats M5 Max in LLM Tests

The benchmark results highlight a key difference in architecture between discrete GPUs and Apple's unified memory. The RTX 5090's raw compute power shines when the model fits within its 24GB VRAM, delivering up to 133% faster Performance in token generation for smaller LLMs. However, as context length grows, the VRAM fills up, tanking performance to 25.28 tokens/second, while the M5 Max's 128GB unified memory keeps it at 181.40 tokens/second.

This trade-off is central to choosing hardware for AI workloads. For users running large models or long contexts, the M5 Max is the practical choice. But for typical local LLM tasks with smaller models, the RTX 5090 offers superior speed, which may justify its cost for enthusiasts. NVIDIA's upcoming RTX Spark, which the article suggests is partly motivated by running heavier AI models, could address this gap, but details are not provided.