Overclocking aint what it used to be

This is sort of my thoughts on something I’ve been mulling over for a while. Love to hear what you guys think too.

Lately, I’ve had the pleasure of dealing with a good bit of older hardware, at least to me! By this I mean my 8120, i7 960, 7970, 290, and a buddies x5650 and 7870. All of these products have performed surprisingly well, especially when overclocking is considered. The differences in GPU and CPU overclocking when compared to today’s offerings is something that stuck me, and something that I feel is worth talking about. Each area has developed and changed differently, and I’ll address each individually. Both are less exciting and less important to overclock now, but for different reasons, and often different depending on the manufacturer. Some are limited by heat, others by other aspects such as design or manufacturing node. However, almost all modern components are limited by some aspect. I want to take some time to talk about my experience, and then look at the typical experience of others in regard to newer and older products, so let’s get into it.

Starting with my 8120, for example, this is a CPU with a base clock speed of 3.1ghz. However, with very little tweaking and effort I’ve managed to have the old chip running at 4.4ghz for over a year now with no issues in stability. That overclock is fairly mild by bulldozer standards, something than 9/10 chips can achieve, across both zambezi and vishera revisions. However, compared to base clock, that is still a hefty improvement of nearly 42%. 42% is nothing to laugh at. Turning to look at the 950, I have a less impressive overclock right now, of 4ghz. Nehalem was never quite as good of a clocker as bulldozer though, so this isn’t too bad, and at these speeds this offers a 30% overclock over the base speed of 3.0ghz. Now, for the big boy, the x5650. These CPU were never intended for general use in consumer PCs, released as a server CPU originally with a MSRP to match at $999. However, as server farms began to sell of old hardware after upgrading, these CPU could be purchased for increasingly lower prices. Now, it is simple enough to find these on ebay for $30-$45. My buddy recently got a great deal on a older x58 system, and wanted to upgrade the CPU, so naturally, a $30 upgrade from a quad core to a hex core CPU made the most sense. With a base clock of 2.66ghz, he is now running at a respectable overclock of 4ghz, or an overclock of over 50%. Needless to say 50% more cores when compared to the typical i7 from x58, and at a 50% overclock for $30 is a crazy value proposition

Moving from my personal experience to common experiences of others, I want to talk about a number of CPU. Starting with the x5650, these westmere chips all seem to overclock like a bat out of hell. There are a number of variations of the 6 core x-series xeons, each with a different max multiplier. Most are able to hit 4ghz no problems, and many are able to hit 4.4ghz or even higher. At 4.4ghz, that is a whopping 65%! That is a MEAN overclock, no matter who you ask. Let’s turn our gaze to bulldozer now, also a great overclocker, if encumbered with other issues. Most bulldozer CPU can hit 4.8, provided you can keep them cool. Given how many different SKUs there are, the level of overclock ranges. Obviously, 4.8ghz on a 8350 is less impressive than the same overclock on a 8120. And again, chips like the 8350 are more likely to hit the clock speeds, however, don’t let that distract you from the fact that many 8120 and similar can handle 4.8, most at 4.6 or higher. So, using the 8350 and 8120 as the points of reference, you’re looking at an overclock of either 20% or 55%. Considering how many of the lower SKU chips can hit these numbers, again that’s one hell of an overclock. Looking at the extreme edition chips, the 5960x could often be brought from the base clock of 3ghz to 4.6ghz, or even 4.8ghz, for an overclock of 53% or higher. Now, I want to talk about another infamous overclocker, the 2500k. This is a chip that even today remains viable and relevant as a gaming CPU, in part due to the massive overclocking potential. With a base clock of 3.3ghz, most users were able to quickly and easily overclock this chip to 4.8ghz, for an overclock of 45%. While not as impressive as the other overclocks, again this a healthy boost for performance. The popularity of the 2500k, along with the massive overclock potential paints a very telling image of overclocking CPUs in days past. Cheap, popular, and already competent CPUs could be pushed much further, for very little, making what was already good great.

So we’ve talked about CPUs, what about GPUs? The performance gains are nowhere near 65% as we saw on CPUs, but the performance gains are still there. My 290 is able to reach clock speeds of around 1180 to 1200, depending on the task. At 1180, that means an overclock of roughly 13%. As for the 7970, is manages an overclock of 1160. Compared to the “ghz edition” 7970, this represents an overclock of roughly 12%. As for my buddies 7870 though, as well as my old 7870, the story starts to change. I’m able to hit 1200 on mine, and his manages 1230. Compared to the base clock of 1ghz, thats a respectable uptick of over 20%. These numbers don’t seem particularly great when compared with the 25% or greater overclocks common to newer pascal cards, but things are interesting when you look at performance. With these older AMD cards, a 5% increase in core speed almost always translates into a 5% increase in performance, this is not the case for pascal. Looking at a few reviews online, it seems that an increase from the base clock of 1486 to around 2ghz nets an increase of only 10-15% in performance. Suddenly, these overclocks seem a lot better. As for vega, we’re looking at a base clock of 1156, with typical overclocks reaching around 1600, sometimes higher, or roughly 38%, a pretty large overclock. However, and this is worth noting for both pascal and vega, boost clocks need to be considered. If we treat the typical boost clock as the true starting point, the margins drop. Vega will typically boost to 1400 or even 1500, pascal will typically boost to around 1700. At these margins, the overclock margins drop from 25% to 18%, and for vega from 38% to 14%. Much closer to inline with what we have been seeing otherwise. However, unlike CPU, my main issue with overclocking isn’t the lower ceiling for clocks, there are other issues at play.

So why am I upset about overclocking now? Starting With CPUs, let’s begin by looking at popular offerings from both sides. AMD is simple enough, the only CPUs that seem to matter on that side are the 1600 and 1700. Intel, on the other hand, is a bit more complicated. The 8700k has completely replaced the 7700k, and to my thinking, there is no reason to purchase one at this point. However, it is still incredibly popular, likely due to users having to upgrade motherboards as well as CPU if they wish to upgrade. Looking at all of these, all ryzen CPU cap out at around 4ghz, some can hit 4.1, and in very rare cases 4.2. Compared to the base clock of 3.2 on the 1600 and 3.0 on the 1700, we’re looking at an overclock of 25% or 33%. Not bad, but not nearly the overclocks we saw before. On the intel side of the fence, the 7700k has a base clock of 4.2ghz, the 8700k offers a base clock of 3.7. Both of these can reliably hit 4.8ghz, with many hitting 5ghz. At 5ghz, over base this means an overclock of 19% on a 7700k, 35% for a 8700k. I believe that here it is worth noting the relatively high boost clocks on intel CPU, which suggests that many of these CPU could likely run at higher clocks with stock settings, but due to the nebulous nature of that claim I won’t consider that here. Simply comparing the 1600 and 7700k to the midrange champs from sandy bridge and bulldozer, we’re looking at overclocks of roughly 20-25% compared to 40-55%, or roughly half the overclocking potential. Granted, the clock speeds are roughly the same in the case of the intel chips, but the overclock potential is much smaller. On the extreme side of things, comparing the modern intel extreme edition flagship, the 7980xe to the 5960x, we see similar overclocking, from a base speed of 2.6 to around 4.4ghz, sometimes as high as 4.8ghz for a total overclock of a whopping 65%. However, this will typically push power draw to the point that only a very high end custom watercooling loop can handle it, limiting the draw.

Looking at GPUs, the story is much the same, but a bit different. Now, we see overclocks of greater magnitude, but the performance gains are roughly the same. All things held constant, there really hasn’t been much of a change. However, there are other things that need to be considered. For starters, hard mods and bios modifications do less for the latest generations than previously. Pascal in particular without heavy modification and changes seems unwilling to go past 2.1ghz, even with top quality components and cooling. When compared to previous cards, where bios mods, voltage increases, and appropriate cooling were the key to achieving higher performance. It seems that no amount of bios tweaking or otherwise can meaningfully raise the limits of either pascal or vega.

On the CPU side of things, the story is fairly similar. Rather than limits seemingly put in place by the manufacturer, there are a number of other issues. Heat is the main enemy in the blue camp. Even with delidding, modern Intel CPUs get incredibly hot, so hot that even with custom watercooling, delidding, and exotic TIMs these CPUs are still limited by heat. On the extreme side of things, power too becomes and issue, with many motherboard VRM designs being unable to keep up with the massive power draw, often overheating due to poor VRM design, or simply not being robust enough. For AMD, due to the design and limits of the 14nm node they use simply cannot move past a certain clock speed limit. These limits have resulted in the same methodology applying for every chip in either generation, with every chip ending up in a very predictable overclock range depending on cooling, which brings me to my final point.

Overclocking used to be more fun. Not only were the performance gains more meaningful, at 45-65% instead of 20-30%, but the method was more fun. Certain chips needed different voltages tweaked, NB overclocking mattered, there were other things to worry about, like bus speed, mem multipliers, qpi, etc. Now, it’s just multiplier, voltage, maybe a little LLC, PLL, or SOC and you’re off the the races. With the hard limit of AMD architecture and Intel temperatures, almost every overclock is a standard 4 or 5, with the same settings tweaked, voltage within a few microvolts of any other chip. Same goes for GPUs, instead of tweaking power and voltage and balancing heat, and running into a voltage or power wall, now there is a wall that seemingly can’t be passed on nvidia, and on AMD the power settings balance against temperature and power draw so much that overclocking without incredible cooling is an impossibility. While the situation on GPUs is much less a problem than on CPUs, it is a problem nonetheless. Gone are the days where tweaking and fiddling with settings could net you that extra little bit, now its a matter of throwing money at the problem to keep something colder, or just simply giving up because the arch wont go faster. Instead of spending the time to learn the system, and tweaking that bus overclock for the little bit better performance over the multiplier (as was the case with bulldozer), now its just 200x25, end of it. Bring back overclocking.


I guess that just means that today’s products are actually good enough.

We are at a point right now where provision of cooling and current to a device from outside drastically change its max clock and performance more so than making timing. As such they can build one product that can span a very wide range of use cases and let the user and AIB decide where to devote resources whether that be VRMs, heat pipes, sinks or water.

This is a good thing. Having the device capable of maximizing its perfroamance at run time in a closed loop manner rather than being gimped by fuses blown or resisters soldered is good.

I’m totally ok with them binning and telling me what they think the max at nominal cooling and current provision will be for a given part and leave me to decide how hard I’m going to work to cool and feed it above and beyond that. It’s making it possible to fit a much wider set of use cases with lower risk of frying parts.

What worries me is being locked and limited because they had to cap at voltage or current they knew their garbage stock heat sink could handle.

yeah, i guess that’s one way to look at it. I’m not upset that the performance is better out of the gate, rather it just seems that overclocking matters less. That’s something I’ve always equated with enthusiasts and building a computer yourself, getting every last bit out of your computer, or getting similar or better performance for diy rather than just slapping something together and only being 10% behind another computer. Just seems that now more than ever throwing money at the problem rather than working for it is just as viable.