One AWS T4G GPU Runs Three Gemma 4 AI Runtimes for Under $3 in Cost Comparison

A developer benchmarked three different serving runtimes — vLLM, JAX, and PyTorch with Transformers — each running Google's Gemma 4 E2B instruction-tuned model on a single AWS g5g.2xlarge instance equipped with an NVIDIA T4G GPU. The experiment used a shared Python benchmark harness across all three deployments to ensure the runtime was the only variable, addressing a prior flaw where each rig measured itself with its own tooling. The entire exercise, spanning 19 instances and roughly four and a half instance-hours, cost less than three dollars in AWS spot compute. The g5g instance family is notable as the only AWS offering that pairs an NVIDIA GPU with an ARM-based Graviton2 host, making it uniquely suited for aarch64 workloads at compute capability 7.5. The project also surfaced five measurement errors that would have gone undetected without a unified harness, underscoring the importance of consistent benchmarking methodology.