Nvidia adds multi-GPU TensorRT inference to Dynamo-Triton

Nvidia released multi-device TensorRT inference support in Dynamo-Triton 26.07, allowing one served model to execute across multiple GPUs behind a single gRPC endpoint. TensorRT 11.0 supplies the distributed capability, while the server manages execution contexts, CUDA streams and NCCL communications for each GPU rank. In Nvidia's eight-GPU Cosmos 3 Nano test, end-to-end video-generation latency fell from 156.595 seconds to 34.183 seconds. The benchmark excluded model loading and MP4 encoding and did not measure concurrent throughput, cost per video or total ownership cost.