Oracle has unveiled a new lineup of GPU compute clusters that will offer AI training services through its cloud infrastructure. The flagship model boasts an impressive configuration of over 100,000 NVIDIA Blackwell GPUs, making it one of the most powerful computing resources available.
This top-tier cluster features a staggering total of 131,072 B200 GPU acceleration cards, achieving peak performance levels of 2.4 zettaFLOPS in FP8 floating-point and INT8 integer computations, equivalent to an astonishing 240 trillion calculations per second. The basic node of this cluster is an NVIDIA GB200 NVL72 liquid-cooled cabinet, each containing 72 GPU acceleration cards. The cabinets are connected through an NVLink bus that provides 129.6 terabytes per second of bandwidth.
In terms of both the number of acceleration cards and peak performance, Oracle’s offering surpasses previous benchmarks set by other industry leaders, including Elon Musk’s innovations. However, it’s important to note that this announcement is still preliminary; Oracle has stated that NVIDIA will not begin bulk shipments of the Blackwell GPUs until the first half of next year, leaving the precise timeline for the deployment of this massive cluster uncertain.
In addition to the flagship model, Oracle is also rolling out two additional clusters. The second cluster is equipped with 16,384 NVIDIA H100 GPUs, achieving a peak performance of 65 petaFLOPS (65 quadrillion calculations per second) and a total bandwidth throughput of 13 petabits per second. The third cluster features 65,536 NVIDIA H200 GPUs, with a peak performance of 260 exaFLOPS (260 quintillion calculations per second) and a bandwidth throughput of 52 petabits per second. This third cluster is expected to become operational later this year.
Companies such as WideLabs and Zoom are already beginning to take advantage of Oracle’s new cluster services.
