AMD Ryzen Overclocking Guide: Get More from Your CPU

[Admin]

Whether you’re using a first-gen AMD Ryzen CPU or the latest Ryzen 3000 series, follow these steps to speed it up.

AMD Ryzen Overclocking Guide: Get More from Your CPU : Read more

 

[Alvar "Miles" Udell]

Long time TH reader, first time commenter.

I have an X370 motherboard, the ASUS Crosshair VI Hero Wi-Fi AC, and I bought a new Ryzen 7 3700X to replace my 1800X. Being an X370 board, things like core sleeping and PBO do not work right, as PBO was only able to reach on 1-2 cores 4.3ghz at 1.45v. Using P-state overclocking, I'm able to get program, game, and AIDA64 stable of 4.3ghz all cores at 1.175v.

So while using PBO may be advisable in some cases, it never hurts to manually overclock and see your results as well. Sacrificing a small bit of 1-2 core performance, assuming I was able to hit the 4.4ghz advertised, for a large power and heat savings while improving all core performance is definitely worth looking into.

 

[JayGau]

Alvar, take a look at this: https://www.pcgamesn.com/amd/ryzen-7-3700x-overclocking-benchmarks-performance

Those guys managed to have something that actually POST at 4.3GHz on all cores with 1.4375V and you claim you do the same and it's stable but with 1.175V? Very hard to believe.

 

[Alvar "Miles" Udell]

There's been a couple of AGESA updates since they did their test, and there's also the silicon lottery. It's not Prime95 stable, but then again neither were my 1800X and FX-8350 and they never gave me any issues with their overclocks in either AIDA64 or normal usage. The voltage does fluctuate quite a bit as to be expected, a bit more than I'd like, +/- 0.025v, but as it's currently over half an hour AIDA64 stressing cache (and counting, see screenshot below), it's stable.

Also remember that it's been several months since the Ryzen 3000 series released, going on a year, giving TSMC time to perfect their yields, whereas PCGamesN used a very early chip.

 

[razor512]

I wish they would do a deep dive into PBO and the clock offsets.
The issue with Ryzen is that there are tradeoffs due to boost states when there are limited thread use.

For example, the R7 3800X will boost to 4.5GHz under lightly threaded workloads, but with a manual overclock, your only choice is an all core OC that may top out at 4.3GHz. Thus a manual overclock will lower your performance in lightly threaded workloads.


Ryzen does not allow for you to tweak the boost curve, e.g., suppose your CPU could handle 4.7GHz on 1 core, 4.6 on 2 cores, and 4.5 on 4 cores, and 4.3 on all cores when at full load. An ideal scenario would be an OC curve that would allow you to enter in a custom curve and the CPU will automatically choose the best aspect of the curve to use based on the workload.

Since that is not available, there is a heavier push to use PBO and the auto OC function, thus we need to find a way to get someone to do a thorough test into it using average cooling such as an NH-D15.

In focusing more on the article, they showed a test run of multiple loops of cinebench. Imagine if they had repeated the test but with better cooling, as well as with some of the PBO settings changed to show us how tyo better maintain our performance. Imagine if they specified the exact cooler used and temperatures for that graph, since precision boost has time limited boost functions that will reduce performance even if temperatures are low, unless you change that multiplier thing in the PBO settings.

While many tasks are multithreaded, there are still make tasks that are lightly threaded, while not ideal, PBO currently offers the best chance at getting a balance of all core clock speed, and maintaining some of the lightly threaded boost speeds, thus appealing to people who are gaming where often there will be a single thread bottleneck, such as the render thread, or the AI or physics thread being a single thread in the game engine, and thus due to a lower clock speed, even though the game is not at 100% load on all cores, you have that physics thread bottlenecking because it is using 1 core worth of CPU time, and everything is waiting on that thread. In cases like that, PBO offers better performance in the game as compared to an all core 4.3GHz overclock. because the bottleneck thread may be at 4.4 or 4.5GHz instead.

Or in productivity tasks, photoshop is heavily reliany on single threaded performance, same with adovbe after efeects and a number of other programs, but then there are effects and tasks which will load all cores equally.

 

[alextheblue]

I wish they would do a deep dive into PBO and the clock offsets.

Agreed. PBO + Auto OC article would be nice. I would like to see if there's any advantages to using Ryzen Master vs just enabling PBO in the BIOS. I'd certainly kick on XMP in the BIOS and make sure FCLK and MCLK are 1:1 (well, as long as you're not pushing your RAM speeds too far, 3600 is the most I can afford anyway).

 

[TJ Hooker]

I actually wish someone would do a follow up investigation on manual Ryzen 3000 OC, with the latest AGESA, and possibly a newly purchased chip. In TH's original review they said the following:
"Our resident overclocking expert Allen 'Splave' Golibersuch has also spent time with early Ryzen 3000 samples and was unable to break the 4.1 GHz barrier without sub-ambient cooling."

A month or so later, the same Splave writes an article where he casually throws out that he overclocked a 3700X to 4.4 GHz all core with only 1.25V, and says:
"In short, fixed frequency IS the fix for now, and maybe forever."

But the seemingly huge change in OC results isn't even acknowledged, let alone explained. I asked him about it in the comments in that second article, but didn't get a response.

It seems like all the info published on manual OC came out with the initial reviews right when the NDA lifted, and any subsequent investigation (as new AGESAs were released or whatever) focused purely on boost behavior, either stock or PBO/AOC.

 

[AMDIs4me]

Alvar, take a look at this: https://www.pcgamesn.com/amd/ryzen-7-3700x-overclocking-benchmarks-performance

Those guys managed to have something that actually posts at 4.3GHz on all cores with 1.4375V and you claim you do the same and it's stable but with 1.175V? Very hard to believe.

I know this is an old message by now, but I just wanted to share my two cents on this. With the latest bios, I can get 4.5GHZ per CCX (CCX0 it seems is my best one) and at 1.425. Now, I would never run my CPU at that voltage for longer than just a benchmark. For the curious ones, I own a MEG X570 (GodLike) currently at 1.3 without trying to fine-tune it one bit, since I am certain my CPU will outlast my need for it before it degrades if it ever does at that voltage.

I can get 4300 all cores on my 3700X, never exceeding 75c under 100% load in my room temp, with a stock cooler. And this is what I currently use for 24/7 along with my ram that could never exceed 3200 without 1.5v lol. I am now at 3800 pushing infinity to 1900 and only 1.4v on Gskill cl 16 (3200) 64GB set.

Previously, I gave up trying to manual OC (when they first came out) since it wouldn't even let me get 4.1-4.2GHZ with anything less than 1.41875

Bottom line: RYZEN OC has come a long way with the latest AGESA updates. And now, this is where the silicon lottery comes into play, I suppose. Some of us can get 4.4 at 1.2 others only at 1.4, like me. 🤷

 

[AMDIs4me]

I wish they would do a deep dive into PBO and the clock offsets.
The issue with Ryzen is that there are tradeoffs due to boost states when there are limited thread use.

For example, the R7 3800X will boost to 4.5GHz under lightly threaded workloads, but with a manual overclock, your only choice is an all core OC that may top out at 4.3GHz. Thus a manual overclock will lower your performance in lightly threaded workloads.


Ryzen does not allow for you to tweak the boost curve, e.g., suppose your CPU could handle 4.7GHz on 1 core, 4.6 on 2 cores, and 4.5 on 4 cores, and 4.3 on all cores when at full load. An ideal scenario would be an OC curve that would allow you to enter in a custom curve and the CPU will automatically choose the best aspect of the curve to use based on the workload.

Since that is not available, there is a heavier push to use PBO and the auto OC function, thus we need to find a way to get someone to do a thorough test into it using average cooling such as an NH-D15.

In focusing more on the article, they showed a test run of multiple loops of cinebench. Imagine if they had repeated the test but with better cooling, as well as with some of the PBO settings changed to show us how tyo better maintain our performance. Imagine if they specified the exact cooler used and temperatures for that graph, since precision boost has time limited boost functions that will reduce performance even if temperatures are low, unless you change that multiplier thing in the PBO settings.

While many tasks are multithreaded, there are still make tasks that are lightly threaded, while not ideal, PBO currently offers the best chance at getting a balance of all core clock speed, and maintaining some of the lightly threaded boost speeds, thus appealing to people who are gaming where often there will be a single thread bottleneck, such as the render thread, or the AI or physics thread being a single thread in the game engine, and thus due to lower clock speed, even though the game is not at 100% load on all cores, you have that physics thread bottlenecking because it is using 1 core worth of CPU time, and everything is waiting on that thread. In cases like that, PBO offers better performance in the game as compared to an all-core 4.3GHz overclock. because the bottleneck thread may be at 4.4 or 4.5GHz instead.

Or in productivity tasks, photoshop is heavily reliant on single-threaded performance, same with adobe after effects and a number of other programs, but then there are effects and tasks which will load all cores equally.

I don't know if you are aware of this, but here is something I personally do. Under Ryzen master, I choose either creator or gaming profile (Creator for me) and then I drag the slider on my fastest cores to as high as it can go, normally I can hit 4.4-4.5 (3700X) with a low enough voltage, especially since it will easily hit those speeds under light load and not even get hammered by voltage. Once you've found the max for your CCX 0 and CCX 1 fastest 4 cores, you can then check under "Task manager--->Details" and locate the software that you are using "set affinity". You can then untick all the cores that are not running at the high frequencies, this will give it access to the so-called fastest cores that are currently running at the highest speed your chip can handle.

This same method could be used for games, as well. I personally don't see it as worthy, and I don't know if you would really gain performance vs all core 4.2-4.3 but I have verified that it works. As for the "Is it better than all core 4.3" I do not know, that is up to the individual to decide... All I do know is that my single thread application(s) will run at 4.5GHZ or 4.7 if your CPU can indeed go that high when I only give them access to said cores through "set affinity". Remember, do not mirror the speed of one core to all the rest. We want to manually boost each of the fastest cores manually, through some trial and error (or HWINFO) you can see how high they can go and get a somewhat reliable voltage level to work with. I personally was able to 4.4 all 4 of those cores with only 1.33125... You see, since those are the only 4 cores that are running fast AF, the other ones are only running at 3520-3600, so the CPU is not requiring high voltage to maintain only 4 cores. Depending on the lottery, of course.