SUPPORT.md currently lists Krackan / npu6 as unmapped. Here is a real-hardware datapoint, not an IREE lock.
Device: PCI 1022:17f0 rev 0x20 (NPU6 / Krackan Point), BDF c3:00.1
Host: Linux (kernel 6.19 class), XRT 2.21.75, /dev/accel/accel0 present, amdxdna loaded, RLIMIT_MEMLOCK unlimited
IRON device: npu2 / aie2p
What ran:
- mlir-aie IRON hello vector-add on NPU: PASS
- Depthwise conv 1D (C=1024, K=9, T=32 and T=375, INT8 activations / INT16 weights) vs NumPy: bit-exact
What this is not:
- Not an IREE
npu6 target lock
- Not a claim that in-tree
amdxdna matches every XRT userspace (we use a version-matched out-of-tree module when needed)
Happy to help fill a npu6 row if you want a specific verify-stack.sh artifact.
Source notes (IRON lab, not a competing IREE pin): WorldTreeNetwork/iron-lab forthcoming; occupancy/doctor work lives in a local tree for now.
SUPPORT.md currently lists Krackan / npu6 as unmapped. Here is a real-hardware datapoint, not an IREE lock.
Device: PCI
1022:17f0rev0x20(NPU6 / Krackan Point), BDFc3:00.1Host: Linux (kernel 6.19 class), XRT 2.21.75,
/dev/accel/accel0present,amdxdnaloaded,RLIMIT_MEMLOCKunlimitedIRON device:
npu2/ aie2pWhat ran:
What this is not:
npu6target lockamdxdnamatches every XRT userspace (we use a version-matched out-of-tree module when needed)Happy to help fill a
npu6row if you want a specificverify-stack.shartifact.Source notes (IRON lab, not a competing IREE pin): WorldTreeNetwork/iron-lab forthcoming; occupancy/doctor work lives in a local tree for now.