Skip to content
luxusAi

lea stability (BIOS)

Phone page for BIOS. Isolation log on the box: ~/.local/share/lea-stability/tests.log.

Stop BIOS A/B. CPB-off + EXPO-off j16 dirty ~4 min (2026-08-25 12:09, load 23, 5325 MHz). The 2026-08-22 CPB-off pass did not reproduce. PPT 90 did not reproduce. EXPO 6000 dies. iGPU off does not help. Daily: keep nix 8/8, do not run kernel -j16.

Now (2026-08-23)

Section titled “Now (2026-08-23)”
Item State
Board / BIOS ASUS ProArt X670E-CREATOR WIFI 3702 (2026-04-09)
Core Performance Boost Disabled (cpufreq.boost=0). Under j16 still 5325 MHz (amd-pstate). Helped only at 4800, not at EXPO 6000.
Precision Boost Overdrive Auto. PPT 90 one pass then confirm fail. Real 120 W dirty ~96 s.
CPU / GPU 9800X3D + RTX 4090
RAM Corsair EXPO II 6000 now. EXPO I/II fail j16 even with CPB off / iGPU BIOS off. Known-good compile was 4800.
iGPU amdgpu blacklisted — one NVIDIA card + renderD128
nix max-jobs=8 cores=8 (j8 passed; seatbelt, not the fix)
memtest86+ 5 passes / ~8 h, zero errors (while EXPO was still on)

Ruled out

Section titled “Ruled out”
  • Thermal — Tctl ~76–80 °C at death, not ~90.
  • 4090 TDP / PSU trip — GPU idle ~18–26 W on compile deaths.
  • EXPO-at-6000 — same kernel -j16 death at JEDEC 4800.
  • iGPU / dual ICD — -j12/-j16 still die with amdgpu unloaded.
  • “any 16-wide gcc” — max (gcc -j16 + 32G RAM + 6× NVENC) passed.
  • Telemetry Pkg is ASUS package temperature, not watts.

Class A cliff (kernel nix rebuild)

Section titled “Class A cliff (kernel nix rebuild)”

Same drv, 4800, iGPU off. Not loadavg (j8 peaked 12.4 and finished; j12 died at load 12.2).

--cores Result
6 pass (daily switch, CUDA ~2 h)
8 pass ~25 min
10 pass ~22 min (2026-08-22; nix-exit 0)
12 dirty reboot ~14 min with boost on; pass ~21 min with CPB Disabled (cpufreq.boost=0)
16 dirty reboot ~2–6 min with boost on (PBO Auto / PBO Disabled / PPT 120); pass ~20 min with CPB Disabled; one pass ~19 min with PPT 90, then confirm #2 dirty ~63 s after EXPO I/II

Compile and games both hammer all-core boost + IMC. Cap nix at 6/8 is a seatbelt. The fix is BIOS/CPU power, then re-test j12.

Chrome/x.com SIGILL is a second class (GPU/compositor). iGPU-off is that experiment — do not mix it with a BIOS PBO change on the same boot if you can avoid it.

BIOS — already learned

Section titled “BIOS — already learned”
  • DRAM 4800, iGPU off, nix 8/8.
  • CPB Disabled → j12+j16 pass. Proof that Precision Boost is the trigger. Not the daily fix (games need boost).
  • CPB Auto + PBO Disabled → j16 still died ~4.5 min. PBO off still allows full Precision Boost 2 at AMD stock PPT. Too little constraint.

PPT 90 is not proven (confirm #2 died)

Section titled “PPT 90 is not proven (confirm #2 died)”

PPT is a package watt budget, not a frequency cap. Logger maxmhz is max(scaling_cur_freq) at 1 Hz — one core at 5.3 GHz still prints 5300. So we cannot see from GHz whether PPT applied.

What we can see:

Boot Profile j16
86963d4d 13:12–13:31 4800, never trained EXPO this cycle, PPT 90 pass ~19 min + nh os switch
EXPO II then EXPO I 6000, PPT assumed 90 die ~65–80 s
b056441e 15:20–15:21 Manual 4800 after EXPO, PPT assumed 90 dirty ~63 s, load 3.5, Tctl 75, 5325 MHz, GPU idle

Confirm #2 died on the ramp, like EXPO, not like the old boost-on 4800 deaths (2–6 min, load ~17). Two remaining causes, one BIOS trip to split them:

  1. PPT reset when Ai tuner left EXPO (Manual/Auto often reverts PBO Limits).
  2. EXPO residue — ASUS Manual after EXPO often keeps DRAM 1.4 V / high VDDSOC / secondary timings at 4800.

PBO Disabled is still not a power cap (stock PPT 162 W). CPB Disabled remains the only profile that passed j16 twice. Not the daily fix (games need boost).

Units (W vs mW)

Section titled “Units (W vs mW)”

PDF has two PPT fields. [W] 90 is 90 watts. [mW] 90 is 0.09 W (SMU ignores → stock 162 W). Check the suffix if you edit it. The old “90 pass / 120 fail ⇒ watts” argument is weaker now that 90 failed confirm.

Menu Item 90 W Eco-ish 120 W Stock 9800X3D
Ai Tweaker PBO (p.21) PPT Limit [W] 90 120 162
AMD Overclocking (p.69) PPT Limit [mW] 90000 120000 162000
AMD CBS SMU (p.45–46) PPT / TDP [mW] 90000 / TDP 88000 PPT 120000 PPT 162000, TDP 120000
AMD CBS SMU TDC_VDDCR_VDD [mA] 90000 110000 (105 W Eco) 120000
AMD CBS SMU EDC [mA] 150000 (65 W Eco) 170000 (105 W Eco) 180000

Look at the [W] vs [mW] suffix on the line you edited. If it says mW and you typed 90, that never applied.

One change: PPT only, correct unit. Leave TDC/EDC Auto. CPB Auto.

  1. Read the suffix on PPT: [W] vs [mW]
  2. [W] → 90. [mW] → 90000. Do not type 120.
  3. If that field is under AMD CBS: PPT Control = Manual first, then the PPT number.
  4. F10. Proof of apply is j16 surviving, not max-MHz dropping (see above).

TDC/EDC only if PPT is confirmed applied and j16 still dies. Then TDC is mA (110000, not 110). The Eco triple (PPT+TDC+EDC) is reference, not this trip.

PBO Enhancement / Eco Mode on 2026 BIOS (not in the 2023 PDF): Ai Tweaker → Precision Boost Overdrive → AMD Eco Mode if that dropdown exists — that is the official cap, not typing 90 into mW.

Core Performance Boost stays Auto.

Stop (2026-08-26)

Section titled “Stop (2026-08-26)”

No BIOS profile survived a second j16. Leave CPB Disabled + DRAM 4800 if that is what is in CMOS; do not chase PPT/Eco/EXPO. Chrome SIGILL on x.com is a second class (renderer, Media thread, no kwin GPU-RESET).