Question Good undervolt?

[Randoman303]

So I decided to undervolt my 5600g and managed to get 30-20-30-30-20 as a stable result, is this good?

 

[Darkbreeze]

As a stable result of WHAT? What exactly are those scores supposed to represent and what were you testing both with and for?

 

[hotaru.hino]

It's hard to say what's "good" if we don't know what you're aiming for either. You could argue any drop in power consumption while retaining the same performance is good, even if that drop is something like 5-10% less.

 

[CountMike]

So I decided to undervolt my 5600g and managed to get 30-20-30-30-20 as a stable result, is this good?

Those must be PBO2 settings per core, How "good" they are depends on particular CPU taking "Silicone lottery" into account. If you have 500 series MB, you can just run Core Optimizer from BIOS in "Per core" mode. If BIOS supports it you can just set negative offset for whole processor.

 

[Randoman303]

As a stable result of WHAT? What exactly are those scores supposed to represent and what were you testing both with and for?

Forgot to add, its a negative per core offset, I tested with prime95 for about an hour and it came back with no errors, and I tested it for lower temps/power usage

 

[Darkbreeze]

Yes, but what are those numbers you posted supposed to represent. I don't mean to be dense, but this is not something I've seen before and I do this every day for the last ten years on this forum, and building/troubleshooting/overclocking systems since the mid 80's.

 

[hotaru.hino]

Yes, but what are those numbers you posted supposed to represent. I don't mean to be dense, but this is not something I've seen before and I do this every day for the last ten years on this forum, and building/troubleshooting/overclocking systems since the mid 80's.

Configuring PBO Curve Optimizer for undervolting only lets you input numbers, which only represents some bucket of sorts. I'm sure you could translate it into a physical value somehow, but basically you're typically limited to -30 to 0.

Though I think OP is missing a number because you can do this per core.

Forgot to add, its a negative per core offset, I tested with prime95 for about an hour and it came back with no errors, and I tested it for lower temps/power usage

If that's the lowest you can go, then sure, it's good.

If you're after getting the lowest values possible, you'd have to win the silicon lottery.

 

[drea.drechsler]

Forgot to add, its a negative per core offset, I tested with prime95 for about an hour and it came back with no errors, and I tested it for lower temps/power usage

Prime 95 isn't a good test of curve optimizer undervolt stability. That's because the undervolt is at the high end of the V/F performance curve. So what's happening is the CPU is using a lower voltage only for the very high frequencies when the processor is boosting, which is when it's handling light, bursty workloads and temp is relatively low. P95 will heat up the processor so it won't be trying to boost to high frequency and so your CO settings are really not being tested.

You can go looking for an application called Core Cycler. It runs some stress apps (you can choose) in a mode that's very light and only on one core at a time. As it's name says, it cycles through all the cores. It's kind of spotty since even though my 5800X passed about 6 full cycles (6 times through all cores) it still would crash sporadically in games. I had to back off the CO settings on some cores to get that straightened out.

But otherwise, I'd say your settings are pretty good; perhaps too good. Most people I've known end up with a -10 to -15 setting for their 'Gold Star' cores. Those cores are already set up to run with lower boost voltage by AMD during manufacture, and so lowering much more usually leads to crashes.

I've read where each increment in the setting co-responds to from 2 to 3 mV in change. Since this is working on a curve it's probably only at the highest achievable boost frequency (5100-5150Mhz for Ryzen CPU's) and scaling downward from there to lower frequencies along the curve. Never seen anything from AMD on that just something in an enthusiast web site (HotHardware.com I think).

 

[Darkbreeze]

Ok, it just totally went over my head what he was referencing with those numbers. It just didn't even click. Sorry for that.

So, as drea.drechsler said, might be a bit TOO good.

I think the first question, now, that ought to be asked is, WHY are you undervolting?

Were you seeing a lack of sustained boost because of having thermal issues?

Were you SEEING thermal issues with the stock or PBO configuration?

Are you manually overclocking or running PBO?

There just isn't much information to go off of from what you posted and basically everything everybody here has said is either a guess or an assumption that you are doing "X" when in fact you might be doing "Y". When you ask about a specific thing, you need to provide ALL of the specifics regarding your configuration, all hardware involved (Including exact model numbers), any relevant conditions, preferably some visual sensor data (HWinfo) and some insights as to what your backstory is for the whole thing. Otherwise, people make assumptions that might not be accurate for your situation. Just my thoughts on that.

 

[hotaru.hino]

I think the first question, now, that ought to be asked is, WHY are you undervolting?

Were you seeing a lack of sustained boost because of having thermal issues?

Were you SEEING thermal issues with the stock or PBO configuration?

Not speaking for the OP, but in my case, I undervolted because I didn't care about absolute performance, I wanted efficiency. Yes shaving off maybe 10-15% power usage (about 10W) during a full core workload might not sound like much, but if it means I don't have to spend as much on cooling or have my cooling setup work as hard on top of the power savings, then that's a win-win anyway. Especially when pushing even more performance out of a modern day CPU barely gets you anything for a lot more power consumption.

 

[CountMike]

In some cases "absolute performance" may be achieved by lowering voltages.

 

[BFG-9000]

Prime95 for an hour does not mean stable. How unstable are you willing to put up with?

I have always ended up with a lower overclock using much higher voltages than everyone else, because I want it to be more stable than stock.

One of the only useful things that Microsoft's telemetry has given us is the surprising realization that PCs have far higher error rates than is commonly believed. Even at stock speeds, the typical whitebox PC (that is, made of aftermarket parts like retail motherboards from ASUS, MSI or Gigabyte) has 3x the error rate of a brand-name OEM PC (like Dell, HP or Lenovo) despite the latter generally being made of far lower quality components like fewer power phases for the CPU, and by someone like ECS (of PCchips infamy). Why is this? It's very simple--the OEMs routinely overvolt the chips to higher than their VID.

More recently, Google and Facebook are finding silent corruption they believe is due to pushing process nodes too small too quickly to develop effective error correction for.

There's a reason AMD binned that chip the way that it did for both speed and voltage. Sure, at lower voltage it may only fail testing when using an instruction that is seldom used, but it failed nevertheless. It's not like in the olden days when there was a lot of unused margin left in the chips that you could use as headroom for overclocking.

 

[Darkbreeze]

Nice, they call it silent corruption now. We've always just called them microerrors, which of course leads to silent corruption. For this reason, and I agree with you that most people would vastly prefer to opt for stability over saving three dollars over the course of a year by undervolting, I pretty much never recommend undervolting. If you are a very advanced user and you absolutely know what you are doing and why you are doing it, then so be it, it's not for me to tell you what to do, but for everybody else I don't think it's a wise decision.

I often hear gamers say things like "I'm not worried about microerrors" because I'm just a gamer and I don't have any mission critical or highly important files on my system. Really? So, when they start getting in game and system errors after a few months we can point to the fact that they didn't feel microerrors creeping into their system and game file data was important, and laugh. In fact, just about every experienced overclocker (And yes, using PBO and other automatic configuration utilities IS the same as overclocking, it's just a different way of doing the same thing.) will tell you that once you find your maximum (Or desired) frequency plateau it is usually wise, if you intend it to be a daily driver type system, to back it off by about 100mhz and bump the core voltage up by whatever minimum increment the board you are working with will allow and the allowable increments are not necessarily always the same for every board. This is, flatly, to ensure (To a reasonable degree anyhow) as much as realistically possible that the system is stable.

Which brings up another point mentioned further up by hotaru.hino where he states he's less worried about absolute performance than he is about efficiency. In which case my response would be, then drop your multiplier by 100mhz and leave the voltage where it is. If it's not manually set and is being automatically configured by the system, it's going to automatically use less core voltage than it would have with a 100mhz higher core frequency and therefore basically be "more efficient" without sacrificing stability to achieve it. Obviously though, opinions are like, well, you know, and everybody has their own so not everybody will agree (With anything, ever, including where the sun rises and whether or not the earth is flat. LOL.) and of course you can do what you wish with your own system. I'm just trying to offer an alternative perspective, or to some degree just second what BFG was saying.

 

[hotaru.hino]

The problem with the paper that BFG linked is it provided no correlation between voltage and stability. It only provided a correlation between clock speed and stability. Also I would argue that system builders who make their own motherboards will have their own testing to ensure relatively higher stability than DIY parts due to the nature of their target market. I find this no different than Intel's "toothpaste" TIM as it performing pretty much as good as any other TIM with the property that it also has a longer endurance rating.

Also with an AMD system, setting a fixed multiplier causes it to no longer automatically adjust its clock speed. Granted there's still some power saving going on because parts of it will be gated when not in use, but you can't set clock speed limits on an AMD system. It's either you get a CPU that can adjust its speed on the fly or you get a fixed clock speed.

Also AMD CPUs have the ability to clock stretch: https://skatterbencher.com/amd-precision-boost-overdrive-2/

Clock Stretching is a safety feature that is built into all AMD Ryzen CPUs. When the CPU thinks the actual voltage is too low to sustain a stable system at a given frequency, it will reduce the clock period until the voltage is back at the acceptable level.

And the whole "saving $3" thing for efficiency only looks at the computer itself and ignores the effects of what having a computer that dissipates more watts can do elsewhere. By finding ways to undervolt or power limit my parts, while finding a sweet spot where I don't lose as much performance, I've shaved off nearly 50-60W total. If that means it's not heating up my room as much during the summer and causing the AC to kick on more (or me wanting the AC to kick on more), then I'm saving even more money that way.

If you guys value stability, then more power to you. But I value other things and in my experience, issues that potentially arisen from instability due to what I've done are so infrequent that I don't really find a need to look for another solution.

 

[Darkbreeze]

If you guys value stability, then more power to you. But I value other things and in my experience, issues that potentially arisen from instability due to what I've done are so infrequent that I don't really find a need to look for another solution.

And honestly, this is what it comes down to. Personal preference. But like I said, you can have it both ways. Stable AND efficient, it just means losing a very small amount of overall performance that 99.9% of people would never be able to identify anyhow short of a very minute difference on synthetic benchmarks.