Ultralytics Benchmarks

Training methodology

We use training throughput — images processed per second during training — as the yardstick; it correlates directly with time-to-solution. Every result is measured on Ultralytics Platform GPUs — the same NVIDIA hardware you rent for cloud training in one click, from entry-level workstation cards up to flagship data-center GPUs. We train YOLO26 at all five sizes (n/s/m/l/x) so you can match a model to your hardware budget.

Settings. 2 epochs on 25% of COCO at 640px, AMP mixed precision, single GPU, with auto-batch (batch=-1) selecting the largest batch that fits in memory. We report the steady-state second epoch, which excludes first-epoch warmup (dataset caching, CUDA graph capture) and the end-of-run validation pass. Resolved batch size, peak VRAM (including the CUDA context), and peak board power are recorded directly from each GPU.

Cost-efficiency. Images per dollar = throughput × 3600 ÷ hourly price, using Ultralytics Platform on-demand pricing — it often reorders the ranking dramatically, as value cards out-earn flagship GPUs per dollar. Measured on ultralytics 8.4.68, torch 2.8, CUDA 12.8. See also the Train and Benchmark mode docs.

Inference methodology

We use inference latency — milliseconds per image at batch 1 — as the yardstick; it's what real-time deployment cares about. Each YOLO26 size is exported to the format and precision shown, then we time steady-state prediction (warmup excluded, per-run times sigma-clipped) on the same Ultralytics Platform GPUs you deploy on.

Formats & precision. On NVIDIA GPUs: PyTorch FP16, ONNX FP16, and TensorRT FP16/INT8 (TensorRT 11.1.0.106 via NVIDIA ModelOpt). INT8 is calibrated on COCO128; accuracy (mAP50-95 on COCO val2017) is measured once per format/precision and shown per row, so the speed↔accuracy trade-off is explicit. INT8 pays off most on larger, compute-bound models — on the smallest models its quantization overhead can match or slightly trail FP16. ONNX inference runs on Ada/Ampere/Hopper GPUs (onnxruntime has no Blackwell kernels yet).

Apple Silicon. Mac devices (e.g. Apple M5 Pro) are profiled locally and merged into the table — pick a format to compare them directly against the cloud GPUs, or filter by the Apple Silicon generation. CoreML runs on the Neural Engine (the fastest Apple path, ~3× CPU), PyTorch on the MPS GPU, and ONNX on the CPU; each row's Backend column names the runtime. CoreML is timed with coremltools on the ANE — a relative host screen, not an iPhone-device number. Owned hardware has no rental price, so cost-efficiency (img/$) is GPU-only.

Cost-efficiency. Inferences per dollar = FPS × 3600 ÷ hourly price, using Ultralytics Platform on-demand pricing. Measured on ultralytics 8.4.71, torch 2.12, CUDA 13.2. See also the Predict and Benchmark mode docs.