Nvidia says its Blackwell GPUs offer massive power and performance gains over Hopper
Nvidia's Blackwell GPUs surpass Hopper in power, efficiency, and performance for various sectors.
Nvidia's introduction of the Blackwell GPU architecture marks a significant milestone in computational progress, delivering massive performance gains over the Hopper GPUs. The Blackwell GPUs, specifically designed for complex calculations required in quantum computing, drug discovery, and other research-intensive fields, promise a dramatic reduction in energy consumption and operational costs. For instance, generative AI tasks on trillion-parameter models are said to be 25x cheaper and less energy-intensive, while weather pattern simulations see a 200x cost reduction and a 300x energy usage decrease.
Notably, the Blackwell GPUs boast a 30% increase in FP64 and FP32 FMA performance over Hopper, showcasing around 45 TFLOPs of FP64 compute performance per B100 GPU. This leap in efficiency and computing power is further amplified in configurations like the GB200 “Superchip,” which pairs Blackwell GPUs with the Grace CPU for even higher performance metrics, such as 90 TFLOPs of FP64 performance. These advancements signify Nvidia's commitment to pushing the boundaries of computational capabilities, catering to the growing demands of AI research and high-performance computing applications.
However, amidst these innovations for enterprise and scientific use cases, Nvidia has remained silent regarding the availability of Blackwell-based GPUs for the gaming market. Speculation suggests that these next-gen gaming GPUs might launch late 2024 or early 2025, but official details are yet to be confirmed. This anticipation underscores the wider interest in Nvidia’s GPU technology, extending beyond professional applications to the gaming community eager for the next leap in graphics performance.