AWS benchmarks Blackwell G7 instances for 30B-model inference

AWS has published SageMaker inference benchmarks comparing G7 instances powered by Nvidia Blackwell GPUs with earlier G5, G6 and G6e families. The tests used Qwen3-Coder-30B for a coding assistant and Nvidia Nemotron-3-Nano-30B for enterprise assistant workloads.

For the coding test, AWS deployed the same model and serving container on each instance and sent 100 synthetic requests at concurrency four, averaging 128 input and 128 output tokens. The tested G7 configuration used two GPUs and 64 GB of aggregate memory, while G5 and G6 each used four GPUs and 96 GB. AWS reports that G7 delivered the highest throughput and lowest latency for this specific setup.

One technical advantage is native support for Nvidia’s NVFP4 format, which stores weights at roughly four bits each. Earlier GPU generations can run those weights but lack Blackwell’s dedicated FP4 acceleration.

The results are workload-specific rather than a universal ranking. Model architecture, quantization, serving software and traffic shape can change the outcome. G7 availability is also limited to AWS’s US East (Ohio) and US West (Oregon) regions, so teams should benchmark their own production pattern before switching.