Skip to content

Section 0 - Getting Set Up

This guide assumes IRON is already installed, as the top-level README describes: the XDNA driver and XRT, pip install -r requirements.txt, and pip install -e . from this checkout.

Check your installation

source /opt/xilinx/xrt/setup.sh
xrt-smi examine            # confirms XRT sees the NPU
pytest iron/tests/operators/test_catalog.py -k AXPY

If that test passes, the toolchain compiles and runs on your attached NPU.

The device

Every operator and graph compiles for the current device, which iron detects automatically from XRT when one is attached. On a host with no NPU (to compile without running, or to record and check a design against its numpy reference), bind one explicitly before anything else runs:

import aie.utils as aie_utils
from aie.iron.device import from_name

aie_utils.set_current_device(from_name("npu2", n_cols=8))

npu1 is Phoenix and Hawk Point (4 columns usable); npu2 is Strix, Strix Halo and Krackan (8 columns). The rest of this guide's examples call iron.operators and iron.Graph the same way whether a real NPU answers or not; only the device run itself (calling a compiled graph) needs one attached.

The cache

Compiled artifacts are cached by content hash under ${NPU_CACHE_HOME:-~/.npu/cache}, nothing is written next to your source. Point a session at its own cache with:

export NPU_CACHE_HOME=/scratch/$USER/iron_cache

The first compile of a design is the slow one; a later compile of the same shape, on the same machine, with the same cache, is near-instant.

Where to go next

Top · Next