NVIDIA RTX Pro 6000 vs NVIDIA DGX Spark (128GB)

Analysis

On paper, the DGX Spark has a compelling LLM inference story: 128 GB of unified memory for roughly $3,000 — more RAM than the RTX Pro 6000 at half the price. The catch is memory bandwidth, which exposes a fundamental architectural difference between the two platforms.

The DGX Spark's LPDDR5X delivers 273 GB/s — less than a fifth of the Pro 6000's 1,344 GB/s. For LLM inference, bandwidth directly sets the ceiling on tokens per second, so even when both platforms can hold the same model, the Pro 6000 will generate tokens roughly five times faster. The Spark's 128 GB ceiling is genuinely useful for loading a 70B model at Q8_0 or experimenting with 405B models at aggressive quantization — workloads where the Pro 6000's 96 GB requires a quant step down. The Spark also runs on an ARM Grace CPU, which requires recompiling some inference software and may not integrate cleanly into x86-centric infrastructure.

Bottom line: The DGX Spark is the better choice for researchers who need to load very large models entirely in memory and prioritize capacity over throughput — running a single 405B model at Q2, or keeping several 70B models simultaneously loaded for evaluation. The RTX Pro 6000 is the better choice for production inference serving where request throughput and latency under concurrent load are the primary metrics. If you're deciding on price alone, the Spark's 128 GB at $3,000 is hard to beat for a memory-bound use case.