OpenVINO™ INT8 vs. PyTorch FP32 baseline
— lower latency
Loading benchmark results…
System under test
Hardware detected by the harness at benchmark time.
Baseline vs. optimized
Memory: RAM, bandwidth and footprint
Installed RAM, measured bandwidth, and how much each model needs to load and run in each configuration.
Kernel profile: what actually executes
Accuracy retained after quantization
Compare two runs
Pick a second run. Side A is whatever the page is showing now; side B can be an earlier run, another machine, or a file you drop on the page.
What it costs per million predictions
Throughput turned into money and energy. Put in what your machine costs to run and the table recomputes for every model and configuration.
What was optimized
Each configuration adds one layer of the Intel® oneAPI stack on top of the unmodified PyTorch model. The weights never change; only how the graph is compiled and executed.
All measurements
No benchmark results found
The dashboard reads results/results.json. Generate it, then serve the repository root and open this page:
python run_benchmarks.py
python -m http.server 8000
# open http://localhost:8000/dashboard/Opening index.html directly from disk also works once results/results.js exists (it is written alongside the JSON).