Dataflow added G4 virtual-machine support with NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, offering 96GB memory and 1.6 TB/s bandwidth.

Google Cloud announced Dataflow support for G4 virtual machines (VMs) powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. The announcement adds this option to Dataflow’s existing support for multiple GPU and TPU types used in AI inference.

Google specifies 96GB of virtual GPU (vGPU) memory and 1.6 terabytes per second of bandwidth for the described GPU offering. These are GPU memory and bandwidth specifications, not total VM system memory, network throughput, or aggregate cluster capacity.

The blog says customers can run inference inside Dataflow jobs using models described as “up to 70B+” parameters. It also identifies continued support for native capabilities including RunInference, right fitting, and GPU-enabled autoscaling, while claiming performance gains over NVIDIA L4 without supplying a numerical comparison in the excerpt.