Tuning and Validation

Change one significant parameter at a time. First establish a reproducible baseline, then find a stable AC value while keeping LLC, DC, frequency policy, and limits unchanged. Memory must already be validated; otherwise, errors cannot be confidently linked to the latest CPU change.

The goal of AC undervolting is the lowest reasonable AC, not the absolute minimum or an undervolt pushed to the floor: balance lower voltage, power consumption, and temperature with full baseline performance, correct computation, and no unwanted CEP intervention. An excessively low AC may produce no WHEA events, BSODs, or crashes, yet CEP may already be reducing Effective Clocks through clock stretching and reducing performance. The absence of WHEA does not prove that the undervolt is optimal.

1. Record the baseline

   Save the profile, BIOS/microcode version, HT, memory settings, limits, temperatures, and results of the selected tests. For a new system, first validate the recommended Intel configuration.

2. Choose operating limits

   PL1/PL2, ICCMAX, and the temperature ceiling must suit the cooling. Do not change them while tuning AC.

3. Choose a moderate LLC

   In my example, this is Level 5. The matching DC values for ASUS are given in 3.2; the entire profile does not transfer automatically. Do not chase minimum Vdroop just for a good-looking graph.

4. Separate AC and DC

   For manual tuning, disable Synch ACDC. Set DC using Table 3.2 and verify AC/DC after boot. Do not use DC to lower reported power.

5. Validate the initial AC

   In this method, after selecting LLC and verifying the applied DC value, setting AC = DC is recommended as a reproducible starting point, not a universal optimum or a value guaranteed to be minimal, safe, or a target. Before lowering AC, validate and record the baseline: test result, Effective Clocks, VOUT/voltage, and the absence of computational errors and signs of CEP intervention. Then lower only AC; DC remains tied to LLC.

6. Lower AC and compare results

   Lower the AC value itself by 0.02-0.05 mΩ per iteration. If errors or reproducible performance loss appear, return to the previous validated value and refine the boundary with smaller steps of 0.01-0.02 mΩ. These numbers are step sizes, not target AC values. Check errors, WHEA, the test result, and Effective Clocks.

   In the original example, AC started at 0.40 mΩ and was then reduced to 0.35-0.30 mΩ. These values illustrate one specific configuration; they do not define a mandatory starting point or a universal lower boundary.

7. Separate the load types

   Test heavy multithreaded operation separately from single-threaded or lightly threaded high boost and load/idle transitions. Lowering AC does not guarantee a solution to a problem at the upper V/F point.

8. Back away from the boundary and repeat

   If an error or performance loss occurs, undo the last step. Keep the final AC above the boundary you found, with a margin confirmed by repeated testing in warm conditions.

y-cruncher screenshot: distinguish Current from Maximum

1.288 V × 244 A ≈ 314.3 W is close to the current Package reading of 313.7 W, but does not prove DC calibration. These are different channels, and their sampling is not guaranteed to be synchronized in hardware. Maxima in different rows may have occurred at different times.

Time alone does not guarantee stability. Record the actual P/E/Ring, workload and limit reasons: passing under limits does not validate the upper ratios. Compare small gains with variation across several baseline runs.

Quick diagnostics

Use Component Stress Tester, not the π benchmark. The screenshot shows the Windows 0.8.7.9547b menu, not a new CPU test. I record the version, binary, threads and memory: the instruction set depends on the binary; AVX2 describes this run. Mem/Thread is memory per thread and varies by component.

First select the required components; then choose 0 - Start Stress-Testing.

WHAT EACH COMPONENT DOES

* The BKT name in Command Lines.txt includes Toom-Cook; the developer states that Toom-3 was removed.