INT8 is an 8-bit integer number format. In AI it is used to store and compute a model's values with less precision than 16- or 32-bit formats, which makes models smaller and often faster, and lets chips do more operations for the same energy.

It is the usual reference precision for TOPS ratings on NPUs, so many headline performance figures describe INT8 work specifically.

Lower precision loses some detail. Converting a model to INT8 can reduce its accuracy, sometimes only slightly and sometimes noticeably, depending on the model and the method. Not every model or layer converts well, and some tasks need higher precision.

When you see INT8 in a spec, remember that it describes a particular kind of arithmetic and may not match the precision your own models use. Test accuracy after conversion on your real data, and compare devices on the same precision and the same model.