Memory planning for edge AI deployments determines whether a model fits on-device without swapping or OOM errors. This tool calculates VRAM and system RAM requirements across quantization levels (FP32, FP16, INT8) and hardware platforms — covering Jetson unified memory constraints, Coral SRAM limits, and Hailo-8 on-chip buffers. Supports NVIDIA Jetson (unified memory), Google Coral (on-chip SRAM), and Hailo-8 / 8L.
{ }
Jetson modules use a unified memory pool shared between CPU, GPU, and OS — there is no dedicated VRAM. On an 8 GB Orin Nano, the OS and runtime consume ~1.5–2 GB before inference begins. Memory planning for edge AI deployments must account for this overhead. Discrete GPU cards (e.g. desktop RTX) maintain separate VRAM, but Jetson has no such separation.
INT8 quantization reduces weight storage 4× versus FP32, but activation memory — which scales with input resolution and batch size — is computed in FP16 even in INT8 networks. This means VRAM and RAM sizing for INT8 models at 640×640 or higher resolution still requires careful accounting of activation buffers, not just weight size.
Module Power Calculator — size PSU and thermal budget alongside memory.
Inference Throughput Estimator — estimate FPS and latency once memory fit is confirmed.
Full Deployment Planner — combine memory, power, and throughput into an end-to-end edge AI BOM.
Jetson modules use a unified memory architecture — there is no separate VRAM. CPU processes, the OS, and GPU inference all share the same physical memory pool. This means your 8 GB Orin Nano isn't 8 GB dedicated to inference; the OS alone uses ~1.5–2 GB.
INT8 reduces weight storage 4× vs FP32. But activation memory — the largest component at high resolutions — is computed in FP16 even in INT8 networks. Runtime activation memory reduction is ~2×, not 4×.
During trtexec export, TensorRT allocates 1–4 GB of temporary workspace for kernel selection and layer fusion. This is a one-time cost at build time — it doesn't consume memory during inference.
±30% for activations and runtime overhead. Weights are exact (calculated from verified parameter counts). Always validate with jtop or tegrastats on device before finalising memory specifications.