HeadroomComponentsNotesDownload
Notes

I tried the three BIOS tweaks everyone recommends. One made it slower, one crashed it, one did nothing.

2026-10-09

Our test PC is a Ryzen 7 7800X3D with 32 GB of DDR5-6400 and an RTX 5080. Its BIOS is full of settings left on Auto, and the internet is full of people saying Auto leaves speed on the table.

So we tried the three that come up most for this chip, one at a time, with the same test before and after each change: Headroom's score, which runs the same fixed work on one core, on every core, and through memory. Here's what happened.

The baseline

Everything on Auto, with the memory's EXPO profile on:

time clock
One core 13.89 s 4.74 GHz
Every core 12.29 s 4.80 GHz
Memory 18.08 s 30.4 GB/s

Curve Optimizer: 25% slower, and it never crashed

The 7800X3D's multiplier is locked, so the usual advice is Precision Boost Overdrive with Curve Optimizer: lower the voltage on every core, and the chip boosts higher for longer. People normally start at -10. This board only offered presets, and the mildest was -20.

stock PBO + curve -20 again
One core 13.89 s 16.30 s 16.15 s
Every core 12.29 s 15.32 s 15.23 s
Clock, every core 4.80 GHz 4.23 GHz 4.31 GHz

17% slower on one core, 25% slower on every core. No crash, no error, no warning. Nothing in Windows would have told you.

The likely reason: the voltage was too low for this particular chip to hold its clock, so instead of crashing it quietly did less work per second. With the curve set back to off, the clocks came straight back to 4.81 GHz.

If we hadn't measured, we'd have believed the undervolt was working. It didn't crash, nothing warned us, and it was a quarter slower.

Precision Boost Overdrive on its own: nothing

With the curve off and PBO left on, one core came in at 13.70 s, inside the normal run-to-run wobble, at the same 4.81 GHz. On this chip PBO bought nothing we could measure.

Faster memory timing: two crashes

On AM5, Auto usually runs the memory controller at half the memory's speed above about 6000 MT/s, and ours did at 6400. Setting it to full speed (UCLK = MEMCLK) is a popular tweak at 6400, because it cuts the time the CPU waits on memory.

On this PC it crashed twice during the test: once a sudden restart, once a blue screen (0x3B, the kind you get when memory hands back bad data). Back on Auto, no crashes. Auto chose half speed at 6400 for a reason.

The fan scare

During these tests the fans slowed down, and before both crashes they stopped. That sounds like a power problem, so we logged everything with HWiNFO during a test.

The power was rock steady: the supply's 12 V line stayed between 12.05 and 12.11 V. What happened was simpler. The CPU went into the test hot, at 80 °C, from what ran before it. Our test draws only about 61 W on this chip, which its cooler handles at 60 to 67 °C, so the CPU cooled down during the test and the fans followed the temperature down. Fans follow heat, not load. The crashes were the memory setting.

What actually made it faster

EXPO. The memory's own speed profile, one switch in the BIOS. When we turned it on (in an earlier post), the same tests on the same PC went from 4800 to 6400 MT/s:

EXPO off (4800) EXPO on (6400)
Memory 20.8 s 18.1 s 13% faster
One core 16.1 s 14.8 s 8% faster
Every core 14.3 s 13.3 s 7% faster

It's the only setting in this whole story that made the PC faster, it's supported by the memory's maker, and a lot of PCs ship with it off.

And one more, for free: don't judge your PC's speed right after a restart. For the first several minutes Windows does its own housekeeping. On our tests it took 4 to 5 cores' worth of the CPU and cost the every-core test about 8%.

So, should you leave it on Auto?

Unless you're an experienced overclocker, yes. Turn on EXPO or XMP, leave the rest on Auto, and if you do try something, measure before and after. The undervolt that made this PC 25% slower looked perfect in every way except the number.

If you are an experienced overclocker

You already know this test wasn't how you'd do it, and you're right. We tried each setting the way most people do: pick the preset, reboot, see what happens. Tuning done properly looks different:

Done that way, there is often real speed to find, even on a 7800X3D. This post isn't arguing with that. It's for everyone else: the people who read "set Curve Optimizer to -20" on a forum, try it, and never find out it made things slower.

Honest note: this is one PC and three settings. We didn't try the Infinity Fabric clock (FCLK), hand-tuned memory timings, PBO's power limits, Curve Optimizer core by core, or anything on the graphics card. And Curve Optimizer results vary chip to chip: plenty of 7800X3Ds hold -20 or more without losing speed. That's exactly why it has to be measured on yours.

Headroom's score runs the same fixed work every time and compares each result with your last one, so a change like these shows up as a number, not a feeling.

Headroom measures this on your own PC, free, and nothing leaves your computer.

Get Headroom