Curve Optimizer and Curve Shaper are how you tune a modern AMD Ryzen chip. Instead of a fixed overclock, you tell each core's boost algorithm (Precision Boost Overdrive) to assume the silicon is better than it is, which pulls voltage down and lets clocks climb within the same power and thermal limits. This guide covers per-core Curve Optimizer, the newer per-frequency Curve Shaper on Zen 5, and how to validate it without fooling yourself.
Curve Optimizer vs. Curve Shaper — what's the difference?
Curve Optimizer (CO) shifts a core's whole voltage/frequency curve up or down by a set number of counts (each count ~3-5mV). Curve Shaper (CS), added with Zen 5 / Ryzen 9000, lets you offset voltage at specific temperature and frequency zones — so you can pull voltage down where the chip is cool and high-clocking, but add a little back where it tends to fall over. CO is the broad stroke; CS is the fine brush.
Phase 1: Hardware Analysis & Constraints
This applies to Ryzen 7000 (Zen 4) and 9000 (Zen 5) on AM5, including the X3D parts. A few platform realities shape everything you do here:
- The boost algorithm is opportunistic. A negative CO doesn't always lower power or temps — the chip often spends the headroom on holding higher clocks longer. That's a feature, but it means "I undervolted and temps didn't drop" is normal, not a failure.
- Every core is different. Your best core might take -30 while your weakest only takes -5. Copying someone else's CO values is pointless; the silicon lottery is per-core here.
- Low stock VID chips take less negative CO. A chip that already boosts at low voltage has less to give back. Don't panic if you can't match a YouTuber's -40 all-core.
- Dual-CCD chips interact. On 9900X/9950X, changing CO on one CCD can shift behavior on the other, so treat each CCD almost like its own CPU and re-check after big changes.
Hard Limits for AM5
VSOC: AMD hard-capped this at 1.30V via AGESA after early Ryzen 7000 burnout failures (chips were hitting 1.45-1.5V SoC). For daily use the community consensus is to stay at or below 1.20–1.25V. Do not exceed 1.30V, ever.
Tjmax: 95°C on both standard and X3D parts. Ryzen is designed to run up to 95°C and will boost into it — that's normal, not dangerous, but a thermal limit (e.g. 85°C) gives you quieter, longer-lived operation.
DDR5 VDIMM: 1.35-1.40V is comfortable daily; 1.40-1.45V is enthusiast territory; 1.5V+ is for benching only.
X3D voltage: the stacked cache is voltage-sensitive. The 9800X3D officially supports overclocking and PBO, but keep CO and any positive offsets conservative on all X3D chips.
Tools You'll Need
| Tool | Purpose |
|---|---|
| HWiNFO64 | Monitor SVI3 core voltage, effective clocks, temps, and PPT/TDC/EDC limit hits |
| CoreCycler | Per-core stability testing (Prime95 SSE / small FFT, y-cruncher) — the standard CO tool |
| y-cruncher | Fast at exposing CO instability; the "Komari" / VT3 loads are popular |
| TestMem5 (Anta777) / Karhu | Memory stability once you start tuning RAM |
| Cinebench R23 + a single-thread test | Before/after performance, single-core and multi-core |
Phase 2: The Tuning Matrix
Step 1 — Baseline and profile your cores
- Update to the latest BIOS (AGESA matters a lot on AM5), load defaults, enable EXPO for your RAM.
- In BIOS: PBO = Advanced, PBO Limits = Motherboard (or set your own PPT/TDC/EDC), and set a thermal limit of 85°C while tuning.
- Record stock single-core and multi-core voltages in HWiNFO (watch SVI3) and your baseline R23 scores. Note which cores are your "best" (CCD's preferred cores boost highest at lowest voltage).
Step 2 — All-core CO to find a safe starting point
Start broad before going per-core. Set an all-core negative offset and walk it back until stable:
PBO : Advanced
PBO Limits : Motherboard
Curve Optimizer : All Core, Negative
Magnitude : start -15 (step 1-2 counts at a time)
Max CPU Boost Clock : +0 to start (add +100/+200 later)
Thermal Limit : 85 COne 9800X3D owner on overclock.net found a deep -35 all-core "felt" great and ran for days — then failed real stability testing and only held at -13 all-core after 12h AIDA cache, Prime95, and multi-hour CoreCycler runs. The lesson: "ran for days without crashing" is not the same as stable.
Step 3 — Go per-core with CoreCycler
Per-core is where the real gains live. Using CoreCycler, test one core at a time and push each core's CO until that specific core errors, then back off by 2-3 counts:
- Run CoreCycler with the Prime95 SSE config (it auto-cycles each core).
- When a core throws an error, note it — that core's CO is too aggressive. Reduce its magnitude.
- Repeat until every core passes. Your best cores will hold much deeper offsets than your worst.
- Then run an all-core load (y-cruncher, R23) because cores stress differently together than alone.
FCLK can buy you more undervolt
Lowering Infinity Fabric slightly (try FCLK 2000 MHz) can increase how much negative CO a chip will take. On AM5, FCLK, memory clock (UCLK) and the memory controller (MCLK) ideally run 1:1:1 — for DDR5-6000 that's FCLK 2000, which is also the sweet spot for most kits. Our DDR5 guide covers the memory side in detail.
Step 4 — Curve Shaper (Zen 5 / Ryzen 9000)
Once your per-core CO is solid, Curve Shaper lets you fix the spots where CO alone struggles. The pattern enthusiasts use:
- Use negative CO to convince the SMU the cores are stronger than stock, improving and sustaining boost.
- Use positive Curve Shaper in the high-frequency or low-temperature zones to add a little voltage back exactly where the chip tends to drop out, without raising voltage everywhere.
- On chips that already run low voltage, you may see little from CO alone — that's when ECLK (BCLK) plus CO plus CS becomes the path to more clock, as several Zen 5 tuners on overclock.net have documented.
Step 5 — Push clocks (optional)
With a stable curve, raise Max CPU Boost Clock Override in +100 to +200 MHz steps and re-validate. The negative curve is what lets those higher bins actually hold voltage-stable. Re-run per-core testing after each bump.
Phase 3: Validation Protocol
CO instability is sneaky. A bad core won't necessarily blue-screen; it'll throw a single calculation error at idle three hours later, or a WHEA, or just silently produce a wrong result. Layer your testing.
| Test | What it proves | Duration |
|---|---|---|
| CoreCycler (Prime95 SSE) | Per-core CO stability — the primary test | 1 full cycle min; overnight ideal |
| CoreCycler (y-cruncher Komari) | Different math, catches what Prime misses | Several hours |
| y-cruncher (all-core) | Combined-load stability | 1–2 hrs |
| AIDA64 (Cache + FPU) | Sustained heat + cache stability | 4–12 hrs for daily cert |
| Prime95 Small FFT | Heaviest thermal/voltage load | 30–60 min |
No single test is "stable"
As gupsterg put it in the source thread: passing y-cruncher doesn't prove there's no issue elsewhere. He spent a day and a half chasing a TM5 failure that only cleared after a timing change — one test passing is necessary, not sufficient. Baseline stable = a clean CoreCycler cycle + 1-2h y-cruncher. Daily stable = add overnight CoreCycler and several hours of AIDA cache/FPU with zero errors and zero WHEA.
Phase 4: Logging Spreadsheet
Per-core CO needs a log or you'll never reproduce a good result. Copy into Sheets/Excel:
| Date | BIOS / AGESA | CCD0 CO (per core) | CCD1 CO (per core) | Max Boost Override | Curve Shaper | VSOC | FCLK | RAM (EXPO) | SVI3 Volt (SC/MC) | Max Temp | Stress Test | Pass/Fail | Notes |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| -20/-25/-15/... | n/a | +0 | off | 1.20 | 2000 | 6000C30 | CoreCycler SSE | baseline curve | |||||
| +200 | HF +2 | 1.20 | 2000 | 6000C30 | y-cruncher 2h | ||||||||
Sources & Credits
Built on community knowledge
This guide draws on the AMD Ryzen Curve Optimizer / Curve Shaper / DDR5 thread on Overclock.net, especially the methodical per-core work shared by gupsterg and the detailed CO logs from contributors like LaFleurr. We verified the VSOC cap, X3D voltage sensitivity, and thermal limits against AMD's current AGESA guidance before publishing. The original thread is worth reading if you want to go deeper on ECLK and Curve Shaper.
Keep Tuning
Curve Optimizer handles the cores; memory is the other half of AM5 performance. Our DDR5 overclocking guide covers EXPO, timings, and the FCLK/UCLK relationship for Ryzen. Running Intel instead? See the Intel 13900K/14900K undervolting guide.
At WebPC Designs we tune Curve Optimizer on every Ryzen build we ship. Want a system that boosts higher and runs cooler out of the box? Reach out for a consultation.