RTX PRO 6000 is an NVIDIA workstation GPU that appears in several recent single-card training and inference projects. NVIDIA used one to run Cosmos3-H-Dreams, a real-time generative surgical simulator; Aurora1.0-150M was trained on 7B tokens with a single RTX PRO 6000; and another author trained a 210M-parameter text-to-image DiT on one such card in 3.5 days.
NVIDIA unveils Cosmos3-H-Dreams surgical generative simulatorNVIDIA released Cosmos3-H-Dreams, built on Cosmos 3 and FlashDreams, learning from clinical cholecystectomy video and robot kinematics to run interactively on one RTX PRO 6000 GPU; NVIDIA calls it the first real-time generative surgical simulator with live kinematic control trained on clinical recordings.
Aurora1.0-150M ReleasedAurora1.0-150M's first-generation model was released on Huggingface; it was trained on 7B tokens using an RTX Pro 6000 and reports performance similar to GPT2‑Small with benchmarks like PIQA 62.24% and Hellaswag 32.20%. An example inference script is provided in the repo.
Training a 210M text-to-image DiT on one GPUThe author trained a 210M-parameter DiT on one RTX PRO 6000 (3.5 days, ~4.2M 256² images) and reports three findings: learned null attention slots dominate cross-attention, flow-matching loss reflects training health not quality, and a training-time timestep shift notably improves sampling quality.
The A.I. Daily