NVIDIA unveiled the RTX PRO 6000 Blackwell professional graphics card in March 2025. This high-end offering stands as NVIDIA's current pinnacle and most expensive product. Overseas, it commands a price tag of around $10,000, while in domestic third-party channels, it's priced between 75,000 and 88,000 yuan.

Built on the Blackwell architecture and fabricated using TSMC's cutting-edge 4nm process, this graphics card packs a punch. It boasts 24,064 CUDA cores, representing a 10.6% increase over the RTX 5090. Memory-wise, it's equipped with a whopping 96GB of GDDR7 ECC memory, triple the capacity of the RTX 5090. Additionally, it features a 512-bit memory interface width and delivers a staggering 1.79TB/s bandwidth. For connectivity, it supports the PCIe 5.0 interface and incorporates multi-GPU point-to-point communication technology.

The core strengths of this graphics card shine in professional domains. Its AI inference performance soars to 4000 TFLOPS, a significant leap from the RTX 5090's 800 TFLOPS. Graphics rendering speed is doubled, and ray tracing performance reaches 380 TFLOPS, outpacing the RTX 5090's 320 TFLOPS. Moreover, it supports confidential computing and MIG technology, enabling the division of a single GPU into four independent instances.

Although benchmark tests reveal only a modest 9%-16% improvement in gaming performance compared to the RTX 5090, the 96GB of memory is a game-changer. It can effortlessly handle the demands of ultra-large-scale model training, such as trillion-parameter model training and climate simulations.

To cater to diverse scenarios, this graphics card is available in three configurations: the workstation version, featuring dual-fan active cooling and a 600W power consumption; the Max-Q version, with a single turbofan and a 300W power consumption; and the server version, which employs passive cooling and supports eight-card parallel operation.