Is an extra 5% of performance worth 10% extra power draw, or given the ongoing energy crisis, would you accept lower performance at lower powers?
Welcome to Esk Computers, I’m Scott and today on the desk of Esk, let’s see how far away from stock clocks we can get in short order, without rummaging around in BIOS screens, and if it’s worth the bother.
It’s that last part that concerns me as we set off on this experiment. I’d say that today’s new CPUs and GPUs do such a good job of managing their own power limits and speeds that they are best left to their own devices unless you are getting into the extreme sports, world record, liquid nitrogen sort of arena.
That’s a stark difference from the old days, where the likes of Intel’s Celeron 300A could pretty reliably go from a stock speed of 300MHz up to 450MHz, and back in those days that pretty reliably meant your computer would be running 50% faster. Whereas today, getting an extra 100MHz out a chip running at nearly 6Ghz would barely be noticeable, even in ideal conditions.
From memory, both the 1080Ti and the 9900K belong to the last generation, or close to at least, where there was meaningful performance left on the table if you ventured away from stock settings, so let’s see if that’s still the case or if these near decade old parts have degraded too much for this sort of behaviour. Also, this is very much the sort of work that will expand to however long you give it, and can even be a hobby by itself, so let’s limit this to one weekend’s worth of occasional fiddling, and also without touching any BIOS settings, as a somewhat pointless challenge.
Let’s look at the 1080Ti first. This is an MSI Gaming X model, which out of the box according to their datasheet runs with a very slight factory overclock, running with an 1683 MHz boost core clock speed, and an actual memory speed of 1378MHz. The effective memory speed is eight times that. I must say that the core clock speed in particular is a lie, it’s always boosted to at least 1900 MHz without any tweaks. Like all 1080Ti’s, it sports 3584 CUDA Units, 11GB of GDDR5X on a 352-bit memory bus, on a PCI Express x16 3.0 slot
For previous videos I did a very slight bit of quick and dirty tuning in MSI Afterburner, adding 250Mhz on the memory clocks and 75Mhz to the core speeds, but we didn’t get into undervolting or optimising anything. Afterburner does make it really easy to just quickly up the clock speeds while running a benchmark and seeing if it sticks, but let’s undo that and get some performance baselines to measure against.
In 3DMark’s Steel Nomad benchmark, we’re getting a score of 2282. In Time Spy, we get a score of 10217.
The Cyberpunk 2077 benchmark at 1080p Ultra settings gets 77.82 average frames per seconds. And Counter-Strike 2 at 1080p Competitive Settings, that being Low but with all dynamic shadows, gets 435.8 average fps.
We’re seeing maximum GPU temperatures of about 70 degrees C, and the system is drawing a maximum of about 470W from the wall.
Now, I expect I’ll wind up in the same place as before, but let’s demonstrate the process. Overclocking more often than not demands that you’ll need to run more power through the device and deal with more heat than at stock. What I’m not proposing we do is the sort of extreme overclocking that requires disabling the on-board protections against the negative effects of this. There is still a risk here, I suppose, but the sort of practically manufacturer sanctioned process we’ll go through today hasn’t gone wrong for me over the years. Still, consider yourself warned, for both the GPU and CPU.
The quickest way of overclocking these on Window is still, I’d say, MSI Afterburner, which doesn’t need and MSI card to work. I think the closest Linux equivalent for this is the utilitarianly named Linux GPU Control Application.
We’ll crank up the Power Limit Slider, which should allow the card to draw more power when it needs to – it’s not constantly cranking up the power by 20% or anything. By default this also pulls up the Temperature Limit, those obviously being dependant on each other but they can be limited separately.
To take a quick first pass at this, I normally start by cranking up Furmark and have it sitting alongside Afterburner, and tweak the clock speeds up by +25MHz at a time, looking out for any glitches in the benchmark. It will eventually crash in some form or another, either with the drivers throwing a wobbler or a complete system hang. Core clocks of +100 Mhz seems too much for it, so we’ll go back to the last non-crashy value of 75 and creep up more slowly. It crashed again at 90, so it looks like +85 might be as far as this goes, in Furmark at least.
Now I’ve done pretty much the same task on the memory clocks, although Furmark’s not the best test for this. We’re doing the same sort of tweaking and again looking out for any weirdness occurring on the furry doughnut. Also pay attention to the fps number it’s giving us, as quite often before it outright crashes or displays artifacting, the performance will actually drop as the memory gets to the edges of its reliability and has to repeat data transfers, so you might need to back of slightly.
I worked my way up to +600Mhz until it first crashed, so again we’ll drop back to +550 and creep upwards, and while 585 seems stable it also seems to drop framerate, so lets’ leave it at +575 MHz. The issue here is that Furmark’s not super demanding, particularly on the memory, so the next steps will depend on how much you care about stability and if you’re going to boast about it on internet forums. We should by rights run an extended soak test to ensure that this is a solid overclock that’s not going to be unstable on extended use.
For this a good methodology would be running a selection of benchmarks and compute tasks. To be honest, I just see if it gets through the 3DMark benches, and whatever game I happen to be playing, and call it a day.
As it turns out, it doesn’t, Steel Nomad crashes, so we’ll do a bit more tweaking downwards and rerunning. My hunch is the memory, which seems wildly optimistic, so I started with that, and after twenty odd minutes of trial and error we got back down to the +250 Mhz memory and +75 Mhz Core overclocks I’d been running at for the previous videos. Funny that, almost like I’ve done this before.
Although it’s still not completely stable. Previously I’d not bothered with moving the power and temperature limits as I’d been running two 1080Tis, and it’s only a 1000W PSU. The max power limit of 117% was causing instability, so I ultimately had to back that all the way off to the stock 100% limit to get the benchmarks running repeatably.
The observant amongst you will have noticed the slider to goose the voltage to the core, but that didn’t result in any more stability so it seems to be drawing all the power it can cope with and the stock limits. We’ll instead look at undervolting later.
After this tweaking we are now seeing a score of 2347 in Steel Nomad, a 2.85% improvement. Time Spy improves by 3.09% to 10533. There’s a 5.47% improvement in Cyberpunk, getting up to 82.09 average fps, and we see a modest 1.79% improvement to 443.6 fps in Counter Strike 2. That’s an average improvement of 3.3%, although the power draw has now increased to 490W, so 4.26% higher. The card’s clock speeds have gone up to closer to 2000MHz.
This is a relatively quick process and I expect it’s getting us almost all of the gains that we’ll be able to achieve in minutes, but if you’re dedicated and can stomach the risk of overvolting and running hotter than is recommended, further optimisations may be possible, but in my experience this process get us 95% of the way there at 5% of the effort. The first pass at getting the overclocks took less than twenty minutes, but the verification runs of the longer benchmarks and tweaking downwards did take another couple of hours, but easily done in an evening.
Moving on the CPU now. The i9 9900K we have here is an 8 core, 16 thread part that out of the box boosts up to 5Ghz on one core, and 4.7GHz on all cores. It has a TDP of 95 W, getting up to 120W while boosting.
Call me undemanding if you will, but when I daily drivered this I didn’t feel it was ever slow enough to go through the bother of trying to boost it further. That’s still my opinion, but, well, that’s kind of against the point of the video. My plan had been to push it harder when I felt it was falling behind, but almost all of my demanding activities were limited by the GPU performance, with the CPU barely troubled.
We’ll again need a baseline, one that’s mostly CPU dependant rather than GPU, so Cinebench 24 seems as good as any. It’s getting 731 multicore at stock speeds. It’s drawing 300 W from the wall while doing this, and is sitting at a comfortable 70 degrees C.
We are at some point going to have to venture into the BIOS screens, but let’s see how far we can get in Windows. This is the Intel eXtreme Tuning Utility, or XTU as it’s incorrectly be-acronymed. We use the older version 7 for the older processors. There’s a newer version for newer processors, and the AMD equivalent is called Ryzen Master, and can perform similar tricks.
Again, this is at your own risk, but the options XTU expose shouldn’t let you do worse than crashing your system. Just be careful you aren’t asking it to apply an overclocking profile automatically on boot until you’re confident it’s stable.
It’s default mode has a one click button that from the tool tips seems to be doing some of what I intended to do manually, and it takes it’s in build benchmark from 3830 to 3875, a 2.5% boost for very little effort indeed. Let’s see if it can be improved in advanced mode.
I suppose we’ll initially target getting up to 5Ghz on all cores when boosting. I remember some people advocating for having it run flat out all the time for the sake of system responsiveness, but that seems like overkill to me. We will however need to sort out the power limits so that it can boost for as long as there is demand to, so we ‘ll change both PLL1 and 2 to 380W.
The sensible way of doing this would be to start with the stock 4.7Ghz all core clock and adding 100Mhz, then having tweaked it upwards see if it can complete a quick Cinebench R20 run – the 2024 version takes too long – then a quick OCCT run.
If that throws up any errors then we could try and apply a bit more voltage to the CPU, which should help with stability. But I suppose the quickest, YOLO way of doing this would be to immediately crank the all core boost to 5Ghz, see if it works. I’m not recommending you do this, but if you aren’t living on the edge you are taking up too much room. I never said I was a role model.
I’ve done so, with a very slight increase to voltage offset, 0.05V, and it gets through it’s in built benchmark, increasing the score to 4088, and it gets a Cinebench R20 score of 4625. So, that’s possibly the goal achieved in very short order, but maybe it can go faster.
Ramping it to 5.1GHz all core, I see that it’s hitting the current limit, which was then throttling the frequency. Increasing the ICCmax to 230A seems to let it boost to 5.1GHz all core for long enough to complete the Cinebench R20 at least, getting 4976, and 4198 in the XTU benchmark so that’s about 9.6% better than the stock settings.
Pushing to 5.3 GHz was a bit much, crashing the system, 5.2Ghz completed once but not consistently.
So back down to 5.1 GHz all core it is, and I ran though Cinebench r23 by accident but it does reveal that after about 5 minutes or so at full tilt throttling will occur, and it appears that’s the motherboard detecting the Voltage Regulator Modules feeding power to the CPU are getting too toasty and dropping the clocks to protect it.
Now, in my real world use cases I’m never going to be running the processor at 100% load for extended use, but if you are, perhaps if you do a lot of rendering or other CPU heavy calculations, you might want to reconsider or compensate somehow. In my setup, perhaps not having a CPU cooler fan blowing down on the VRM is maybe a factor, and more heatsinks or strategically positioned fans might remedy this.
But, again, life on the edge. Full send it is, and oddly Cinebench 2024 mostly behaves itself better, and now gives us a score of 752. Which is, to be honest, slightly disappointing, getting just about 3% better performance for a 16.67% worse power draw of 350W and temperatures getting up to 90 degrees. In my opinion, a bit spicy for continued use, but again, nothing I do regularly hammers the CPU at 100% for extended periods so it’s maybe only a theoretical concern.
For example, it’s running much cooler in when going through the graphics focused benchmarks again, as it’s not hitting 100% load. Steel Nomad oddly drops a little to 2333 compared to the last set of GPU overclocked results, a 0.6% reduction, although pretty much margin of error stuff. Time Spy makes up for it, up by 1.04% to 10643. Cyberpunk 2077 goes up 0.78% to 82.73 average fps. It’s Counter Strike 2 that benefits the most from the faster processor, going up 4.26% to 462.5 average fps.
On average, that’s an additional 1.37% of performance, albeit for an additional 6.12% or 30W of power draw, taking it up to 520W pulled from the wall. That’s 10% more than the stock settings power draw, for 4.71% better average performance.
There’s a full science behind getting the absolute maximum out of your processor, which will require checking how consistent the voltage delivery to the CPU is, determined by your motherboard and power supply, and also to a degree if your processor is still behaving as well as it was when it was brought into this world. Silicon does degrade over time and these parts are knocking on for a decade old at this point.
However that will mean a trip to the BIOS settings, which is against my self-imposed rule set for this video. Why? Well today I don’t feel like messing around with load line settings and vdroop measurements, and in truth I don’t quite understand them well enough to spout off in a YouTube video about them. Also, it just takes too long, with so many reboots, tweaking, testing, reboot, tweaking, repeat, that it can take days. I’m just trying to show off what’s possible in a shorter time frame.
It is also possible to overclock memory beyond manufacturer settings, which would also require a journey to the BIOS screens again. I’ve never had much luck pushing beyond the manufacturer’s factory XMP settings so I’m going to stick with them. Of course, if you’ve not enabled that in the BIOS and it’s stuck on the default JEDEC speeds and timings, make sure that’s rectified. I know I said no BIOS tweaking, but I’m just showing you this for the visuals. I didn’t change anything. The ruleset remains valid. I remain valid. I am strong and powerful and deserve to be loved. Hang on, what were we talking about? Oh, yes, overclocking.
The traditional way of summing this up would be to say we’ve got nearly 5% better performance for free, averaged over all today’s benchmarks but it’s not exactly for free. The increased power draw is a thing, and according to Statista the UK has the third highest electricity prices in the world. We’re paying for that literally with the extra 10% plus of power draw, and warmer running equipment that means louder, noisier fans.
Maybe instead of peak performance, we should be optimising for efficiency.
Let’s see how these components respond to undervolting and underclocking, that is to say giving them less juice and seeing what they can squeeze. Now, for me, this is the rabbit hole we could vanish down indefinitely. With overclocking, you’ll quickly bump into the voltage limits, power draw and temperature capacity of your CPU, GPU, motherboard or cooling solution, and that’s going to define how far you can push upwards.
Going the other way, you’ll find a practically infinite number of duty point combinations of lower clocks, voltage and power draw that might or might not satisfy your needs, so I’ll just show a really brief first run through what can be done with the GPU and CPU, in the interests of keeping this to the self-imposed weekend’s time limit, not uncoincidentally the time I have available to produce this video.
Here we are back in MSI Afterburner, this is the curve used for undervolting a card. It determines what voltage is supplied to the card when it reaches a given clock speed. I say undervolting, you could equally well provide more voltage at a given higher clock speed if that helps with stability, as long as it’s not over the card’s power limit, and you can deal with the additional heat produced.
I’m just going to pull this entire curve down by about 100 mV, with the max clocks coming down concordantly. You’ll want to look up the keyboard shortcuts for moving these points about, moving each one individually with the mouse is a mug’s game.
The CPU overclock is still in place, but let’s see what it’s done to the benchmark results. The GPU clock speed drops to a bit over 1800 MHz, and the Steel Nomad result drops to 2178, that’s about 4.5% less than the stock speeds. Time Spy gets 10076, that’s only 1.4% less than stock. Cyberpunk gets 77 average fps, down 1% from stock. Counter Strike 2 actually remains up by 6% over stock at 461.3 fps, being carried by the faster CPU speeds. The power draw at this point is still up at 490W, again pulled up by the CPU overclock, so let’s sort that.
Back in the XTU, I’ve come up with a slightly whacky scheme of keeping a two core boost of 5.1 GHz, but when using all cores, it’s only going to get up to 4.5 GHz. So, both a overclock and an underclock. Let’s feed this to Cinebench 2024. Under load, this drops the temperatures to 70 deg C rather than the 90 deg C of the balls out solution. We get a multi-core result of 690, so that’s 5.6% down from the stock settings, but the power draw drops about 23% to 230W from the wall in this CPU test. This seems like a trade-off worth consideration, at least.
Let’s do a final run through of the graphics benchmarks and see if the CPU’s further hurt them. Steel Nomad scores essentially the same 2179 result, so still 4.5% down from stock, and compared to the maximum overclocked results of today’s test, down 6.6%.
Time Spy scores 9844, that’s down 3.6% from stock, and 7.5% worse than the max overclock result.
Cyberpunk 2077 technically improves slightly with 77.45 average fps, but that’s still 0.5% less than stock settings and 6.4% worse than the overclocked results.
Finally Counter Strike 2 takes quite the hit, dropping to 418.4 fps, that’s 4% worse than stock settings and 9.5% worse than the overclocked results. But still, complaining about over 400fps seems churlish.
The average performance drop from stock settings over these tests is a not insignificant 7.5%, but the maximum power draw drops down to 430 Watts from the wall. That’s a decent 8.5% power savings from the stock settings, and a massive 17.3% drop from the overclocked power draw.
I think these results show it’s at least considering using less power rather than obsessing over more. The actual performance loss in games might not be too significant, and cheaper electricity bills are always welcome. Still, if overclocking does push some game or activity over your threshold of playability, a bit more on the power bill is a lot cheaper than a new graphics card, particularly in the ongoing hellscape of 2026.
As for me, while it’s always tempting to spend a lot of time tweaking the setups to get it running as fast as possible on as low a power draw as possible, I’m going to return the CPU to the stock settings and keep the GPU overclock in place, so basically back where we were at the start of this video. That’s partially for ongoing testing for other videos, and partially for the ease of use for the next user when I eventually sell this on. If this were still my daily driver, I think I’d be leaving it underclocked on all cores and overclocked on one or two cores, which I think is the best of both worlds.
As always, if you have any questions or want further details please leave a comment down below, and if you enjoyed this videotronic missive then consider subscribing. Until next time, take care of yourself, and each other.