A recent analysis compares the Nvidia RTX PRO 6000 (Blackwell architecture) with a setup of four RTX 5090 GPUs, showing that the PRO model can match their performance in AI inference tasks.
The test focused on running the MiniMax M2.7 AI model, which has 230 billion parameters (230B). Using a GGUF quantization method (IQ3_XXS) to fit the model within available VRAM, the results were very close between both setups.
RTX PRO 6000 Matches Four RTX 5090s While Using Far Less Power
The RTX PRO 6000 reached 118.74 tokens per second (tok/s), while the system with four RTX 5090 GPUs achieved 120.54 tok/s working together.
However, the biggest difference appears in power consumption. The setup with four RTX 5090 GPUs reached peaks of around 2,300W, while the RTX PRO 6000 used only 600W. This is expected since it is a single GPU instead of four, but it still shows a major advantage in efficiency.
Lower power use also means lower operating costs and less need for complex cooling systems, which is important in professional environments.
From a cost point of view, the RTX PRO 6000 is also competitive. Its price is around $9,500, while buying four RTX 5090 GPUs would cost about $14,000.
This comparison shows that even though consumer GPUs like the RTX 50 series are very powerful, professional GPUs can perform better in some cases thanks to better optimization and no need for communication between multiple GPUs. This gives the Blackwell PRO architecture a clear advantage in AI tasks like inference.
