NVIDIA admits to only 2x performance boost at max throughput with new generation of Rubin GPUs

NVIDIA has provided a more grounded performance expectation for its next-generation Rubin AI GPUs, revealing that the R200 chip will deliver only about a 2x throughput increase over the current flagship Blackwell B200 in maximum production workloads. This admission comes as the company moves to more direct "apples-to-apples" comparisons, contrasting chips with similar memory configurations rather than the previous practice of comparing an 80GB HBM3e Blackwell chip against a 288GB HBM3e Rubin chip. The 2x figure represents the real-world gain for the vast majority of companies running data centers at full utilization.

Despite the Rubin architecture's impressive theoretical specs—including nearly 3x the memory bandwidth and approximately 5x the FP4 performance—these improvements translate to a more modest doubling of output. This performance uplift comes with a significant power cost: the R200 has a Thermal Design Power (TDP) of 2300W, compared to the B200's 1000W. This means data centers would use 2.3x more power per GPU to achieve 2x the performance, raising immediate questions about the generation's power efficiency and total cost of ownership for large-scale AI training and inference clusters.