NVIDIA RTX PRO 6000chat_bubble_outlineFrom real comments

The 96GB of VRAM is the card's defining feature for AI workloads. Multiple commenters highlight that the huge VRAM pool lets them run massive local LLMs, such as 120B parameter Q4 models, entirely in VRAM without RAM bottlenecking, and enables large-scale simulations like 1.8 billion grid cell CFD runs. One lab reported deploying eight of these cards in an LLM research server, calling them awesome. This capacity also supports heavy 3D rendering workloads that are otherwise VRAM-starved on consumer cards.

Bandwidth advantage makes it dramatically faster than NVIDIA's own DGX Spark. Commenters who own both devices state there's no situation where the DGX Spark outperforms the RTX PRO 6000, pointing to over five times the memory bandwidth (1.8TB/s vs 273GB/s shared) as the reason for far higher inference speed. They call claims of Spark superiority disingenuous, valid only for edge cases needing more than 96GB but under 128GB VRAM.

Impressive PCB engineering comes with driver and price trade-offs. Reviewers repeatedly marvel at the tightly packed, dense PCB design, calling it insane and among the most powerful dense boards ever seen. However, some note the card lacks game-ready drivers unlike the 5090, though a workaround exists via NVCleanstall. Several commenters also flag the price as steep, with one estimating a $300 per die markup over the 5090's VRAM cost.