News

[enthropy7]

It achieved ORT results on low-cost ARM processors—the kind listed in my benchmarks, such as the OPI Zero 3. On very low-cost Cortex processors, the results rival what XNNPACK achieved using hand-written assembly, feature gating, and unsafe C. All of this takes on a whole new dimension considering that the framework doesn’t require you to integrate any obscure backends like in Python, but works on the principle of “write the code, compile it, and it runs fast.” All that’s left is to work with the thread pool, tune large inputs, run more models, and assemble a cluster of boards for testing.