Published October 25, 2026 - Updated October 25, 2026 - By Xilly
How To Undervolt Your Ryzen CPU With Curve Optimizer (2026 Guide)
Set up PBO and Curve Optimizer on any Ryzen 5000, 7000 or 9000 CPU, tune it in Windows without rebooting, catch clock stretching, then prove it stable with OCCT and CoreCycler.
BIOS Optimization Guides PC Optimization Guides
Precision Boost Overdrive (PBO) and Curve Optimizer are the two BIOS settings that get the most free performance out of a modern Ryzen CPU. This guide works across Ryzen 5000, 7000 and 9000 series. The example system is a dual-CCD Ryzen 9 9950X3D, so it covers tuning each CCD separately. On a single-CCD chip like the 9800X3D or 7800X3D the steps are the same, you just tune all cores at once.
The testing method has also changed since our last Curve Optimizer video. A lot of people went all the way to -40, never crashed, and never saw their score go up. Below is exactly why that happens and how to test properly so it does not happen to you.
Stock versus PBO + Curve Optimizer combined. Ryzen 9 9950X3D in Cinebench R23 multi-core: 42,361 to 45,324 (+7.0%). Ryzen 7 9800X3D in the Black Ops 7 benchmark, CPU 1% lows: 422 to 451 FPS. Your gains depend on your chip, cooler and motherboard.
What PBO and Curve Optimizer actually do
PBO lets the CPU push harder when it has the room to. AMD's default limits are built to be safe on cheap coolers, so with a decent cooler you are leaving performance on the table.
Curve Optimizer is an undervolt: the same clock speed with less voltage. Less voltage means less heat, and less heat lets the CPU boost higher and hold that boost for longer. That is why it is close to free performance.
Step 1: Update your BIOS and download the tools
Update your BIOS first. Newer BIOS versions carry AGESA fixes that change how boost and Curve Optimizer behave. If you have not done it before, our BIOS update video is linked below. Then download these programs:
- BenchMate: Cinebench R23, HWiNFO and other benchmarks in one place.
- SMU Debug Tool: changes Curve Optimizer inside Windows, so you do not have to reboot for every test. It needs PawnIO.
- OCCT: the stability test.
- CoreCycler (optional): a per-core overnight test.
- How to update your BIOS (video guide).
Step 2: Get your baseline score in Cinebench R23
Before you touch the BIOS, record a starting score. Close everything first: Discord, browsers, game launchers, all of it. Do this before every single test in this guide.
- In BenchMate, open HWiNFO64, choose Sensors Only and click Start.
- Scroll to CPU (Tctl/Tdie) and CPU Die temperatures and keep them visible on a second monitor or the side of your screen.
- Open Cinebench R23. Under File, check Advanced Benchmark and set Minimum Test Duration to Off. That makes it a benchmark rather than a stress test.
- Run the multi-core test, watch your temperatures while it runs, and write down the score.
Why some people hit -40 and never crashed
Watch CPU (Tctl/Tdie) during that first run. The temperature limit is 95°C on most Ryzen 7000 and 9000 CPUs and 89°C on the 7800X3D. Our 9950X3D sat around 75°C. If yours shoots straight to the limit, your cooler is holding you back, not your voltage.
A CPU sitting at its temperature limit cannot boost, and a CPU that cannot boost cannot crash from an undervolt either. That is how some people reached -40 with zero errors: the chip was never boosting high enough to expose the problem. Curve Optimizer will still help a thermally limited CPU, but Cinebench will not show it. Fix cooling first if you can.
Step 3: BIOS settings for PBO and Curve Optimizer
Restart and press Delete or F2 to enter the BIOS. Switch to Advanced Mode if you are in EZ Mode.
ASUS motherboards
Go to Ai Tweaker, then Precision Boost Overdrive.
MSI, Gigabyte, ASRock and other motherboards
Go to Advanced, then AMD Overclocking (accept the warning), then Precision Boost Overdrive.
Scalar: people worry that 10X degrades the chip. We have run 10X on AM5 for months without issues. On AM4 we would stay at 2X or Auto. If you are unsure, 2X is the comfortable choice.
Boost override: +200 MHz and the undervolt fight each other, so the more boost you add, the less you can undervolt. Add it if you are tuning for gaming. If you only want lower voltage, heat and power draw, leave it disabled.
Curve Optimizer starting values
Scroll down to Curve Optimizer and set the sign to Negative. Start everyone at a magnitude of -15. For a lot of CPUs, -15 is honestly the limit, and that is fine: it is still a solid undervolt with a real performance gain. Do not chase a big number. The goal is a curve that gives you zero issues and actually performs better.
| CPU | Curve Optimizer mode | Starting value |
|---|---|---|
| Single CCD: 9800X3D, 7800X3D, Ryzen 5, Ryzen 7 | All Cores, Negative | -15 |
| Ryzen 9 9900X / 9950X (non-X3D) | All Cores, Negative | -15 |
| Dual CCD X3D: 9950X3D, 7950X3D, 7900X3D | Per CCD, Negative | CCD0 -15 / CCD1 -10 |
Why the split on dual-CCD X3D chips? The two CCDs are not the same. CCD0 has the 3D V-Cache and is where your games run. CCD1 is the frequency CCD: it already runs faster out of the box, so it cannot undervolt as far. On a 9900X or 9950X both CCDs are frequency CCDs, so All Cores works fine and per-CCD tuning is not worth the extra time.
Global C-State Control: Disabled
In the Advanced menu (usually under AMD CBS), set Global C-State Control to Disabled. When the CPU wakes from a deep sleep state it ramps voltage up quickly, and with an undervolt it sometimes cannot keep up, which causes random freezes at idle. Disabling it prevents that and gives slightly lower latency. Idle temperatures rise a little, which is nothing to worry about with a proper cooler.
Some people say this breaks boosting on chips like the 7800X3D. That has not been our experience: across hundreds of 7800X3D client systems, they still boost to 5050 MHz. There are reports online of it causing stutters, so test it on your own PC. Then save and exit.
If your PC only stays stable at idle with C-states disabled, your Curve Optimizer may still be too aggressive. The daily-use test at the end of this guide is what confirms it.
Step 4: Tune in Windows with SMU Debug Tool
Back in Windows, open SMU Debug Tool as administrator (install PawnIO if it asks). On the PBO tab you should see the values you set in the BIOS: on a dual-CCD chip, the first group of cores is CCD0 at -15 and the second is CCD1 at -10.
Run Cinebench before changing anything. Just the starting values put our 9950X3D almost 4,000 points over its stock baseline.
The method: find your stable limit
- Go 5 further: -15 becomes -20. Click Apply.
- Close everything, then run Cinebench twice. Let the CPU cool for a bit between runs so heat does not drag the score down.
- If the score went up and there were no errors, go another 5 further.
- Keep going until Cinebench errors out, the PC freezes or you get a blue screen.
- Back off a few steps. If you errored at -25, settle at -22 or even -20.
On a single-CCD chip, select all cores and change them together. On a dual-CCD chip, tune one CCD at a time: do CCD0 first and leave CCD1 alone. If you change both and it crashes, you will not know which one caused it.
How our 9950X3D tuning went
| Step | What happened | Decision |
|---|---|---|
| CCD0 -20 | No errors, same score (run-to-run variance) | Keep going |
| CCD0 -25 | Score up slightly, then 200 points lower on the second run | Probably not stable |
| CCD0 -30 | No crash, no score gain | Clock stretching |
| CCD0 -22 | Highest score so far | Final for CCD0 |
| CCD1 -15 | About 300 points gained, held on a second run | Keep going |
| CCD1 -20 | 200 points up, then back to the -15 score after cooling | No real gain |
| CCD1 -30, then -25 | No meaningful difference | Settle at -15 |
Final curve: CCD0 -22, CCD1 -15. Dual-CCD chips can be harder to crash in Cinebench because the load spreads across 16 cores, so instability shows up randomly. On a single-CCD CPU an unstable value usually errors out fast.
Clock stretching: why lower is not always better
If you push Curve Optimizer too far, the CPU protects itself. The reported clock speed stays the same, but the actual work it does goes down. That is clock stretching. The PC looks completely fine and does not crash, but you are getting less performance. That is why you watch the score, not just whether it crashed.
You can see it in HWiNFO: when Core Effective Clock drops below Core Clock under load, the chip is stretching.
If the score goes down or stays flat two steps in a row, you went too far. Go back to the value that gave your best score. If -25 and -40 give the same score, run -25: same performance, far more stable. Cinebench tells you if it is faster. OCCT tells you if it is stable.
Step 5: Lock your numbers into the BIOS
SMU Debug Tool is your rough draft. Once you have values that give your best score without clock stretching or errors, enter them in the BIOS. Applied from the BIOS, the CPU trains slightly differently, so a value that passed in Windows can crash or score lower and you may need to back off a little more.
Restart, enter the BIOS, and go back to Curve Optimizer (ASUS: Ai Tweaker → Precision Boost Overdrive → Curve Optimizer; other boards: Advanced → AMD Overclocking → Precision Boost Overdrive → Curve Optimizer). Set the values you found, for us CCD0 -22 and CCD1 -15, then save and exit and run Cinebench again.
Expect temperatures to climb a little compared to stock. With lower voltage and the +200 MHz override, the CPU has the headroom to boost higher, and HWiNFO will show it holding higher clocks. If the score matches what you got in SMU Debug Tool, move on.
Step 6: Stress test with OCCT
- Open OCCT and go to the Stability Test.
- Choose the CPU test, set the mode to Variable and the intensity to Normal. Leave everything else on Auto.
- Click Start, wait the 10 seconds, then click Start again.
- Run it for at least 15 minutes with zero errors.
Variable load keeps changing, which is exactly where an unstable undervolt falls apart. If the curve is too far, OCCT usually catches it in the first few minutes. If it shows an error, or the PC freezes or restarts, back off by 5 (for example -22 to -18) and run it again.
This matters even more on dual-CCD chips: games barely touch CCD1, but OCCT loads every core. And if you are coming from our last video already sitting at -30 or -40, go straight to OCCT. It will probably fail almost instantly. If it does, back off by 10 and keep backing off until it passes.
Step 7: CoreCycler (optional overnight test)
CoreCycler tests one core at a time and is the most thorough check for Curve Optimizer. Run the .bat file. If it asks for the .NET Desktop Runtime, Ctrl+click the link, install it and relaunch. When it asks to add a custom event log view, choose Yes: some instability shows up as WHEA hardware errors in the event log without causing a crash.
Each core runs for 6 minutes per iteration, so one pass on a 16-core 9950X3D takes about 96 minutes, and roughly 50 minutes on an 8-core CPU. Run it overnight or while you are away. One clean iteration is a good minimum if you want to be fully sure.
Step 8: The real test is daily use
Curve Optimizer rarely crashes under full load. It crashes when the PC is doing nothing, or jumping from idle to load. So after the stress tests:
- Play your games for a few hours.
- Leave the PC idle for a couple of hours.
- Do a full shutdown and cold boot the next day.
If you come back to a frozen PC, or it restarted on its own, back off your curve. No crashes in games, no freezes at idle and clean cold boots mean you are stable.
What you get
The CPU runs cooler, boosts higher and holds it longer. In games, the biggest difference is in 1% lows, which is what actually makes a game feel smoother: our 9800X3D went from 422 to 451 FPS CPU 1% lows in Black Ops 7. Lower voltage also means less wear on the CPU over time.
When you are done, uninstall SMU Debug Tool and PawnIO. BenchMate, OCCT and CoreCycler are worth keeping for later checks. For the rest of the BIOS on an X3D chip, see our Ryzen 7 9800X3D BIOS guide, and on a dual-CCD X3D make sure games stay on the V-Cache CCD.
Next: Ryzen 7 9800X3D BIOS optimization · Keep games on the right CCD on a 9950X3D · X3D Gaming Mode: should you enable it?
Frequently Asked Questions
What Curve Optimizer value should I start with?
Start at -15 on All Cores. On dual-CCD X3D chips like the 9950X3D, 7950X3D and 7900X3D, start CCD0 at -15 and CCD1 at -10. Then move 5 at a time, testing each step in Cinebench.
Why did -40 never crash on my CPU?
Usually because the CPU was sitting at its temperature limit and could not boost high enough to expose the instability, or because it was clock stretching. Check your temperatures in HWiNFO and compare Core Effective Clock with Core Clock, then confirm with OCCT.
What is clock stretching?
When the undervolt goes too far, the CPU keeps reporting the same clock speed but does less work per clock to protect itself. Your benchmark score stays flat or drops even though nothing crashes.
Is PBO scalar 10X safe?
On AM5 we have run 10X for months without issues. On AM4 (Ryzen 5000) we recommend 2X or Auto. If you want the most conservative setting on any platform, use 2X.
Does CPU Boost Clock Override work on the 7800X3D?
No. It does nothing on the 7800X3D, so leave it on Auto or Disabled there.
Should I tune per CCD on a 9950X?
Not usually. On the non-X3D 9900X and 9950X both CCDs are frequency CCDs, so All Cores tuning works fine. Per-CCD tuning matters on dual-CCD X3D chips, where the V-Cache CCD and the frequency CCD behave differently.
How long should I stress test Curve Optimizer?
At least 15 minutes of OCCT CPU Variable with zero errors, ideally one CoreCycler iteration overnight, then a few days of normal use including idle time and cold boots.
Professional PC Optimization
Want This Done For You?
BIOS, Curve Optimizer, RAM and Windows, all tuned by hand for your exact PC and tested for stability before the session ends.