AMD ZEN 3 PBO & CURVE OPTIMIZER OVERCLOCKING GUIDE

​

DISCLAIMER

​

• By unlocking PBO limits you are violating AMD’s stock configuration and therefore invalidating your Warranty
• Even though this guide is aimed at everyone, I am expecting you to at least know some of the basics about how ZEN cpus work, this includes PBO, PBO limits, navigating BIOS, troubleshooting potential issues that arise, etc.
• Some of the things in this guide will vary from CPU to CPU due to but not only, silicon quality variation, cpu SKU (5600, 5800, 5900, 5950X), cooling method used, RAM setup, Operating System bloat, etc.

​

SOFTWARE

​

• HWINFO64 (https://www.hwinfo.com/download) - Monitor temperatures, clock speed, voltages, etc.
• CPU-Z (https://www.cpuid.com/downloads/cpu-z/cpu-z_1.98-en.exe) - Quick and dirty benchmark for single and multi-thread performance
• OCCT (https://www.ocbase.com/) - All in one stability testing tool, very good, also support the developer, really nice guy
• CoreCycler (https://www.overclock.net/threads/single-core-prime95-test-script-for-zen-3-curve-offset-tuning.1777112/) - Very decent tool to complement OCCT, to test each core individually. Props to blu3dragon from OCN for this tool.
• Ryzen Master (https://www.amd.com/en/technologies/ryzen-master) - Tool to monitor % of TDC and EDC values during testing.
• Other software to validate performance gains such as Cinebench R20/23, 3D Mark suite, Geekbench,gaming benchmarks, etc, can be also used.

​

PRECISION BOOST OVERRIDE aka PBO

​

• PBO ADVANCED

Inside your BIOS, enable PBO and select PBO advanced, this will bring up a bunch of options:

• PBO LIMITS

The value for these limits varies hugely from CPU to CPU, some CPUs scale differently, specially with TDC and EDC combo. Also, SKU matters, the values for a 5600X are absolutely not the same as the ones for a 5950X,

There’s 2 approaches to these limits and I will share the approach that is more user friendly but not the one that will necessarily yield better performance. Further testing for those who want can be done.

Load up BIOS defaults, go into PBO menu and enable advanced. In the advance section of PBO, set PBO limits to motherboard or manual and set values that you won’t realistically hit. Once you do this, boot into Windows, open Ryzen Master and start CB23 multi thread test. Observe TDC, EDC and PPT values and check what % of the max you are hitting. This should be a good starting point as the values to pick for PPT, TDC and EDC.

For people who want to go further, you should play with TDC and EDC combo for higher results, even a small variation can be enough to squeeze a bit more performance.

• PPT (W)

200W is enough for 5600, 5800 and maybe 5900X SKUs. For the 5950X this value is very important because given the chance your CPU will not hesitate going there given the workload. Cooling here is very important because not many cooling solutions will effectively cool a 5950X at 250W. My advice for 5950X users is to use a value between 200 and 300W and test accordingly to your type of workloads.

• TDC (A)

Somewhere between 90 to 150A on 5600, 5800 and 5900X. For 5950X, between 140 to 220A. Test accordingly in CB23 because even a small variation of 5A might bring big gains in multithreaded performance. CPU-Z also a good way to quickly measure performance changes, but it’s not as sensible as CB23.

• EDC (A)

Somewhere between 120 to 200A on 5600 5800 and 5900X. For 5950X, between 140 to 220A. Test accordingly in CB23 because even a small variation of 5A might bring big gains in multithreaded performance. CPU-Z also a good way to quickly measure performance changes, but it’s not as sensible as CB23.

​

• PBO SCALLAR

Change this to x1. This way you assure PBO will not try to override the FIT controller into using a higher level of voltage for longer.

​

CURVE OPTIMISER

This is where all the magic happens, really. This is the single best tool AMD has provided Zen 3 users with. This is the tool that makes the guide come together into a very beautiful thing.

What Curve Optimiser does is apply a voltage offset, positive or negative, to each individual (or not) core’s VID. Basically, AMD CPUs (and Intel and any other CPUs but we’re focusing on AMD here) use a standard “fit all” CPU voltage/frequency curve because individually binning each CPU would take forever and would not be cost efficient. What Curve Optimiser lets us do is tune this curve ourselves so that even the crappiest CPU can take advantage of lower operating voltages and temperatures while increasing performance.

Anyway, testing… The boring part but the most crucial. I prefer to do individual core testing. For this, load up PBO, Advanced, and go to Curve Optimiser. Inside Curve Optimiser, select per core. In this menu you will see your cores, select negative on each of them.

Normally people will tell you best cores do less undervolting and worse cores do more undervolting and while this is true, we cannot forget Curve Optimiser offsets are an order of magnitude and not an actual value. Just because a core does -30 and another -25 it does not mean that -30 > -20 in absolute terms because the core that is at -20 might already be requesting lower VID to begin with.

Either way, we can start by setting each core at -10. Now what I would suggest you to do is to either use OCCT or CoreCyler. I prefer CoreCycler myself.

• OCCT

In OCCT, select Test, CPU, Data Set - Large, Mode - Extreme, Load Type - Variable, Instruction Set - AVX2. In the threads section you can select advanced, physical only, select all cores, and on core cycler section, select cycle active core every 5 minutes.

This ensures you test every core with cooldown intervals between them while sort of simulating what would go on during a game or similar workload where load keeps switching between cores.

Alternatively you can run SSE instruction set and medium to small data set. This will better simulate a gaming load I believe.

​

• CORECYCLER

Pretty straight forward, once you set it up, run it and leave it running. It will automatically keep note of the cores that failed and will automatically skip them for the next tests. Leave it running for the whole duration for faster testing. Do not stop just because a core failed.

​

• TROUBLESHOOTING

Obviously, some cores will fail and some will pass. If the cores pass, you can go -5 (so if you’re at -10, you go -15), for the ones that failed, depending on how fast they failed on CoreCycler (1st, 2nd or 3rd test), I would reduce accordingly. If it failed on 1st test, it means the core simply cannot handle that undervolt. So back off 5, if it fails on 2nd or 3rd test, you can back off 3 or 2 (so if you’re at -10 you go -5, or -7 or -8). For OCCT, I don’t think there’s a cause/effect where you can deduce how bad a core is, I guess if it fails fast it’s bad…

Hard reboot? Don’t know why? Was idling and crashed? Don’t worry, Windows has a beautiful tool to help us determine what core is giving us issues. Go here and check this guide I made about troubleshooting (https://docs.google.com/spreadsheets/d/1SiLpWVL4-T3vdHZKPA2TELPKa7TbJyCGF_JJdjsHdLg/edit#gid=1831618223)

Another tip, from my experience, bad cores (use HWINFO for this) will usually undervolt a lot, we’re talking -20 to -30, while fast cores will be usually below -10. This can help you speed up the testing process.

AFTER ALL OF THIS IS DONE, BACK OFF -1 OR -2 ON EVERY CORE TO ENSURE MAXIMUM STABILITY.

​

FREQUENCY OVERRIDE

This value goes from 0 to 200 Mhz since AGESA 1.1.0.9. whereas previously it would go up to 500 Mhz on MSI and ASUS boards. This value basically tells PBO to try and boost as high as it possibly can. Too high and you get clock stretching, too low and you leave performance on the table.

I usually recommend going straight to 200 Mhz. Keep in mind that this value is hugely tied to curve optimiser, without it, you’ll be leaving a lot of performance on the table. Also, the maximum will probably only be achieved by your 1 or 2 best cores and only by very small periods of time. If you have good cooling (big AIO or custom loop), sustaining this during CB23 Single Thread test is actually possible. CPU-Z single thread is a very fast and somewhat reliable test to check for changes in single core performance. For this, simply select the thread box and chose 1. This will only use 1 core and you can affectively measure 1T performance.

• DISCLAIMER: CPU-Z uses Core 0 by default for it’s 1T benchmark so if Core 0 isn’t your best core, it’s natural you won’t see as big of a gain, however, it’s still there. To get around this load CPU-Z on your best core and try again.

GENERAL NOTES

• Do not set manual Vcore voltages
• Do not change stock/auto LLC (Load Line Calibration)
• Do not change Scalar from x1.
• Cooling is very important, PBO scales with temperatures, after 50C you lose Mhz for each degree you climb. Good AIOs or Custom Loops are pretty much essential for someone who wants to milk the last bit of performance.
• RAM tuning is similarly if not more important for Ryzen CPUs than PBO and CO tweaking. I would strongly advise everyone and their mother to read this insane guide by fellow members of the OC discord server. (https://github.com/integralfx/MemTestHelper/blob/oc-guide/DDR4 OC Guide.md). As an eample, I tested SOTR benchmark between 3600c16 XMP, 3600c16 tunned and 3800c14 tunned setup and gained over 40FPS AVERAGE on my own setup. Seriously, the gains are ridiculous, much more than this. Games that are very CPU bound such as Call of Duty Warzone will see INSANE gains... I cannot stress this enough, a 6700XT is enough to max that game out graphically, don't listen to people on 3090's with 100 FPS... It's totally CPU bound. Tune your RAM, tune your CPU and you will see insane gains on most games that are CPU bound (RTS, MMO's, MMORPGs, etc.)

ADITIONAL STUFF

Wouldn't be an overclocking guide without some test results right?

Here's my own 5800X on various benchmarks:;

CPU-Z - https://valid.x86.fr/v6k4aw 702 ST - 7072 MT

Geekbench 5 - https://browser.geekbench.com/v5/cpu/6488736 / https://browser.geekbench.com/v5/cpu/6451542 - 1841 ST - 12270 MT (one of the fastest Zen 3 CPU on normal cooling)

CB23 - My PR is 16800 MT and 1690 ST, usually hoover around 16500 (https://cdn.discordapp.com/attachments/802676130741223437/903756463875424288/2541314.jpg)

TS CPU Score - my PR is 14000 , usually hoover around 13800 area (https://www.3dmark.com/spy/22201612)

CPU Profiler on 3D mark - https://www.3dmark.com/cpu/75741 (one of the fastest scores under normal cooling)

​