PC Watch tests 64GB RTX Spark laptop: runs 27B LLM locally, but bandwidth limits speed
PC Watch benchmarked an ASUS laptop with NVIDIA RTX Spark and 64GB unified memory, showing it can run 27B parameter LLMs and generate short videos locally, though its ~300GB/s memory bandwidth results in significantly slower inference than high-end desktop GPUs.
Japanese tech outlet PC Watch has published detailed benchmark results for the ASUS ProArt P16 laptop, equipped with the NVIDIA RTX Spark N1X chip and 64GB of unified memory. The tests aim to evaluate the practical capabilities and limitations of running local artificial intelligence workloads on mobile hardware.
For Large Language Model (LLM) inference, the test used the Qwen3.8-27B model (Q4_K_M quantization), achieving an average generation speed of approximately 32 tokens per second. This rate supports real-time text interaction and code assistance. However, when the input context length increased to 65,536 tokens, the time to first token reached about 102 seconds, indicating a significant latency bottleneck for long-document processing.
In image and video generation tasks, the device required approximately 14 to 40 seconds to generate a 1024x1024 pixel image and about 69 seconds to produce a 5-second video using ComfyUI workflows. While these times allow creators to iterate on visual concepts on-site, they lag behind high-end desktop GPUs.
The primary cause of this performance gap is memory bandwidth. The RTX Spark N1X offers a maximum memory bandwidth of approximately 300GB/s, compared to 1,792GB/s for the NVIDIA GeForce RTX 5090 desktop GPU used as a comparison. Although RTX Spark provides larger model hosting capacity through shared CPU/GPU unified memory, its data throughput rate constrains the execution speed of compute-intensive tasks.
Additionally, the tests noted that not all 64GB of memory is available for GPU acceleration, as the operating system and applications require resources. In Windows settings, the memory allocation for the GPU was adjusted to a cap of 40GB. This underscores the need to distinguish between 'total memory capacity' and 'effective VRAM equivalent' when planning local AI deployments.