You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
examine: Strix Halo APU (gfx1151, Ryzen AI Max+ 395) misclassified as discrete GPU (has_apu=false) #449
On an AMD Ryzen AI Max+ 395 (Strix Halo, gfx1151), rocm examine --json classifies the machine as having no APU ("has_apu": false, GPU entry "is_apu": false, "has_discrete_amd": true). This is a discrete-GPU-only misclassification of an iGPU-with-unified-memory APU, and it feeds wrong routing to consumers that branch on has_apu (APU-specific diagnostics and GTT/memory-tuning flows get skipped on exactly the machines where GTT tuning is the whole point).
Unified-memory evidence: mem_info_vram_total=8589934592 alongside a 96 GiB GTT (ttm.pages_limit=25165824) — classic UMA APU split
Expected
has_apu / is_apu should be true for Strix Halo and other Radeon-A APUs.
Suggested fix
PCI-topology heuristics alone miss APUs whose iGPU enumerates on a separate host bridge (Strix Halo exposes the GPU at 0000:66:00.0). Consider classifying by one of:
gfx target set: gfx1150 / gfx1151 / gfx1152 are APU-only LLVM targets — a whitelist here is cheap and precise
UMA detection: VRAM carve-out is DRAM-backed (no separate HBM/GDDR instance) — amd-smi already exposes this shape
Note: gfx_target being empty in examine output (see #393) is likely also feeding this — fixing the gfx-target lookup would give this classification the cheap precise signal.
Impact
Downstream consumers branching on has_apu misroute: e.g. the official rocm-doctor skill's APU flows and GTT-related diagnostics are skipped on this machine even though GTT tuning (currently 8 GiB carve-out + 96 GiB GTT) is this platform's headline feature.
Ground-truth outputs from other official tools on the same machine, backing the issue body. One local directory-name segment in env.PATH-derived text is redacted.
Additional internal inconsistency in the same rocm examine --json run: the per-GPU entry has "gfx_target": "", while the top-level summary on the very same run reports "detected_gfx_target": "gfx1151" — two probe paths inside one invocation disagree, and the empty per-GPU field is what starves APU classification (related: #393).
Activity
slamdunk111 commentedon Sep 26, 2026
Ground-truth outputs from other official tools on the same machine, backing the issue body. One local directory-name segment in
env.PATH-derived text is redacted.Additional internal inconsistency in the same
rocm examine --jsonrun: the per-GPU entry has"gfx_target": "", while the top-level summary on the very same run reports"detected_gfx_target": "gfx1151"— two probe paths inside one invocation disagree, and the empty per-GPU field is what starves APU classification (related: #393).Excerpt:
and from the same JSON:
"summary.detected_gfx_target": "gfx1151".sysfs / cmdline one-liners (same machine, this boot):
amd-smi static (full) — MARKET_NAME Radeon 8060S, TARGET_GRAPHICS_VERSION gfx1151, MEM_CARVEOUT present
rocminfo (full) — GPU agent Name: gfx1151, Marketing Name: AMD Radeon 8060S Graphics
fix(examine): classify an APU from the kernel target when PCI cannot …
fix(examine): drive the planted host, and keep an undecoded target un…
fix(examine): classify an APU from the kernel target when PCI cannot …
fix(examine): drive the planted host, and keep an undecoded target un…
docs(examine): say which path reads the 1586 row, and what the guard …
fix(examine): classify an APU from the kernel target when PCI cannot …
fix(examine): drive the planted host, and keep an undecoded target un…
docs(examine): say which path reads the 1586 row, and what the guard …
fix(examine): classify an APU from the kernel target when PCI cannot …
fix(examine): drive the planted host, and keep an undecoded target un…
docs(examine): say which path reads the 1586 row, and what the guard …
fix(examine): classify an APU from the kernel target when PCI cannot …
fix(examine): drive the planted host, and keep an undecoded target un…
docs(examine): say which path reads the 1586 row, and what the guard …
docs(examine): say which path reads the 1586 row, and what the guard …
fix(examine): classify an APU from the kernel target when PCI cannot …
fix(examine): drive the planted host, and keep an undecoded target un…
docs(examine): say which path reads the 1586 row, and what the guard …