NVIDIA has unveiled cost details for its Blackwell GPU, highlighting a significant shift in cost efficiency for AI inference tasks. While the Blackwell GPU's unit price is nearly double that of its predecessor, the Hopper, at $2.65 per hour compared to $1.41, it offers a dramatic reduction in cost per token. The Blackwell GPU achieves a 65x increase in token throughput, reducing the cost per million tokens from $4.20 to $0.12, assuming all software optimizations are enabled.
The analysis, based on the DeepSeek-R1 model, shows that with optimizations like FP4 low-precision inference and Multi-Token Prediction (MTP), the cost per million tokens can drop to as low as $0.11. Without MTP, the cost is approximately $2.35 per million tokens. These figures underscore the potential for significant cost savings in AI inference tasks using the Blackwell GPU, although results may vary with different models and configurations.
NVIDIA's Blackwell GPU Doubles Price, Slashes Token Cost by 35x
免責事項: Phemexニュースで提供されるコンテンツは、あくまで情報提供を目的としたものであり、第三者の記事から取得した情報の正確性・完全性・信頼性について保証するものではありません。本コンテンツは金融または投資の助言を目的としたものではなく、投資に関する最終判断はご自身での調査と、信頼できる専門家への相談を踏まえて行ってください。