Introduction
AMD Ryzen 9 7950X "Zen 4" is the fastest desktop processor you can buy right now. When our reviews of the Ryzen 7000 series went live last week, the buzz among tech forums and social media was surrounding its 95°C operating temperature, which the processor was found not just hitting quite often, but basically residing in, under nearly all serious processing loads. AMD made several statements around this temperature, the most notable of which, is that one shouldn't be alarmed with the 95°C load temperatures if the processor's basic cooling requirements are met, and should expect this to be its normal load temperature residency.
95°C is a mighty high temperature, especially for a processor built on the 5 nm node—the most advanced ever for an x86 processor. In their bid to outdo each other, both Intel and AMD have thrown traditional notions of "high" thermals and power-draw out of the window. Intel's flagship Core i9 "K" processors come with 125 W base power marked on the tin, but actual power limits set as high at 241 W for the 12th Gen Alder Lake, expected to go up to 253 W with Raptor Lake. AMD, on the other hand, has raised its TDP (base power) for its Ryzen 9 7000 series to 170 W on the box, but with a power limit of 230 W. AMD is sticking with the classic definition of TDP, and they set the cooler recommendations accordingly—at least a 240 mm AIO liquid cooler is recommended for the Ryzen 9 7000 series. But what if you gave these processors cooling below this recommendation? Will bad things happen? What about air-cooling?
Even with weaker cooling, the processor won't bug out with thermal-trips or emergency system shutdowns, provided you have some form of cooling. It will still hold on to the 95°C temperature at the highest possible clock speed. This in and of itself isn't new. When any chip hits its temperature limit, it either shuts down, or dramatically drops voltages and clock speeds to attempt to lower temperature; only that's not what is happening with Zen 4. The 95°C temperature doesn't necessarily cause the processor to freak out in any way. Instead 95°C is what the processor is trying to run at, by picking the highest possible boost frequency that doesn't cause the temperature to go over the 95-degree mark. Remember, this isn't a classic TJMax, but a load-temperature limit at which the processor begins to lower its highest boost frequency based on the cooling performance. Once you enable overclocking mode, the 95°C temperature target gets disabled and the CPU may run at up to 105°C, and only above that it will turn off automatically to protect itself—this is the real TJMax. We asked AMD, and they confirmed that at stock settings, increasing the 95°C temperature target isn't possible, you can only lower it.
In this article, we are running a Ryzen 9 7950X processor with cooling solutions below AMD's recommendation of a 240 mm AIO. This includes the Noctua NH-U14S, a highly capable air cooler with heatpipes; and the AMD Wraith Spire, a very basic in-box cooler that's rated for just 95 W TDP. Just to clarify, this is not the boxed cooler of the 7950X—AMD does not include a Wraith Spire (or any cooler) with the 7950X. The idea here is to see how the processor behaves in a "sub-par" cooling environment, whether it's capable of stable 24/7 use with air-cooling, just how much its boost headroom is affected with all this, and what its impact on performance is.
Test Setup
All testing in this article is performed using the Ryzen 9 7950X, which is the most powerful Zen 4 CPU available, with the highest heat output. AMD is very clear that the recommended cooling configuration is a "240-280 mm AIO", but we wanted to know what you really need and how it affects temperatures and performance. The rest of the platform is identical to the one in our Ryzen 9 7950X review.
The following cooling solutions were used in this article:
- Arctic Liquid Freezer II 420 mm: this is the AIO that gets us the lowest temperatures, thanks to its Ryzen offset mounting mode, which moves the center of the cold plate away from the middle of the IHS, to sit on top of the compute dies instead.
- Noctua NH-U14S: a typical "high-performance" air cooler.
- AMD Wraith Spire: this cooler is bundled with various older AMD processors, and serves as example of an affordable mid-range cooler without heatpipes, or other advanced tech.
- In order to simulate other cooling solutions from different vendors, with various capabilities, we're adjusting the fan speed on these two coolers using the motherboard's fan control. To clarify, the fan speed is set to a fixed percentage, there's no temperature-dependent regulation, to make it a fair comparison.
We're testing the following application workloads, which have been selected to cover a wide range of the power usage spectrum, at various threading levels:
- Blender (235 W): a highly parallelized rendering application that fully loads all cores. This is the highest power consumption, higher than even Cinebench.
- Cinebench single-thread (44 W): one of the most popular "quick and easy" benchmarking tests. In this scenario it's limited to running on a single-core only.
- Cinebench multi-threaded (216 W): the benchmark that everyone uses these days. It uses the Cinema4D rendering engine, using as many cores and threads as the processor has available.
- Visual Studio C++ Compile (90 W): testing a processor with just a highly parallelized software is like testing a car full throttle only. The vast majority of workloads today are actually mixed threaded, so sometimes they use a few threads, then they drop to one or two threads, and then spin up more again. That's the reason why our compiler test is part of this article.
- Adobe Photoshop (67 W): the software that no creative mind can live without. Adobe products are not highly optimized to scale across all threads, even though some filters are optimized for parallelization. We run several images through a typical Photoshop workflow of various tasks and filters.
- MP3 Encode (43 W): converting raw audio into highly compressed MP3, with minimal loss in quality is what powers all listening experiences today. This benchmark is purely single-threaded and serves to investigate real-life performance with just a single thread (I consider Cinebench 1T a "synthetic" test).
For gaming, we're running:
- Counter-Strike Global Offensive (61 W @ 1080p): a highly popular multiplayer title that's very light in its graphics requirements and thus runs CPU limited all the time.
- Cyberpunk 2077 (90 W @ 1080p): the engine is designed to use as many cores as possible and spread its workload accordingly. Highly GPU limited.
- DOOM Eternal (112 W @ 1080p): this game has the highest CPU power consumption in our tests, but is also heavily GPU bound.
- Far Cry 6 (77 W @1080p): Ubisoft's engine is very compute-intensive, but doesn't scale to all threads; it also puts lots of stress on the memory and inter-thread communication.
Temperatures
There's a huge difference between gaming temperatures and rendering, which isn't that surprising. No game can put as much stress on the CPU as highly parallelized workload—and most "normal" applications don't use as much power either.
With the AIO, temperatures are slightly below the 95°C temperature target that Ryzen 7000 operates at (it's a bit hard to see in the chart). The Noctua cooler will always run at 95°C in rendering, but gaming temperature gradually goes up as we adjust the fan speed. At 20% fan speed, the Noctua reaches 95°C in both gaming and applications.
The AMD Wraith Spire cooler reaches 95°C in both gaming and rendering, even at 100% fan speed, so I didn't include the other RPM settings, which would be 95°C/95°C, too. Great news, even with seriously weak cooling your processor will not overheat and not exceed 95°C.
If you wonder "how can the temperature always be 95°C, even at lower fan speed?" That's how Zen 4 is designed to operate. The CPU will aggressively boost clocks and voltage as long as it's below 95°C, and once it reaches that point it will carefully regulate the boosting to always stay at 95°C (the algorithm's target is "at", not above or below).
Frequency Scaling
Obviously, if the temps are always kept at around 95°C, even with vastly different cooling performance available, then the processor must run at different clock speeds, depending on the cooling solution. With weaker cooling, more heat will build up inside the CPU, so the clock speeds will go down to lower the heat output, to achieve thermal equilibrium around 95°C.
To test this, we're running a workload at different thread counts and observe how the CPU frequency varies with cooling. The frequency reported is the average of all threads that are loaded, i.e. idle cores sitting at 800 MHz are not part of the averaging formula.
As you can see, even the Noctua running at full speed sees a little bit of loss in frequency compared to the AIO. With lower fan speeds and cooling capability, the CPU clocks drop—very gradually. An interesting result here is that even the single-threaded frequencies go down, even though they are not even close to 95°C on any of these coolers. Zen 4 doesn't just lower clocks the moment it hits 95°C, but rather it adjusts the frequencies to the cooling configuration—at all thread counts, even at lower temperatures.
It's also impressive to see that the loss in frequency isn't that big, there's no "falling off a cliff", except for maybe the Wraith Spire running at 40% and lower, with the CPU fully loaded on all 16 cores / 32 threads. All other tests are running at above 4 GHz, even with the Wraith Spire at 20%, which really provides VERY little cooling performance at that setting.
Here's the clock frequency results as table, to get a better feel for the numeric values. Given the round "5750" and "5500" numbers on the AIO, I suspect that AMD has also added some sort of artificial frequency cap to their processor, above which it will never go, no matter how good the cooler.
Performance Scaling
First we're comparing against the Noctua NH-U14S. To make it easier to get a feel for the differences, the AIO performance results are marked as "100%", and the Noctua results are relative to that.
In applications, nearly all workloads see some small performance drops with weaker cooling. Interestingly, the single-threaded MP3 encoding workload takes a noteworthy hit when the Noctua is running at just 20%, comparable to what we're seeing on other multi-threaded tests like Blender and Cinebench. This confirms our findings from the clock frequency test. Certainly the most important takeaway is that the performance differences are really small, even 2.5% performance loss on average will be something you'd never notice. Props again to AMD for making the thermal throttling so well-behaved and gradual.
For gaming, there's barely any differences, no matter the resolution. The only exception is Far Cry 6, which loses up to 5% in FPS at 720p, but the GPU-bound 4K setting runs at the same FPS, even with the Noctua slowed down to just 20%.
Remember, we're adjusting the fan sped on the Noctua, to simulate weaker coolers from other manufacturers, and this is great news, because it means that any half-decent tower-style cooler will be able to run the 7950X (and all weaker Zen 4 CPUs, too), with good performance. Especially in gaming the differences will be negligible.
I also tried running with just the Noctua heatsink, i.e. the fan stopped completely. With this config the machine would bluescreen and crash very often.
Next, we're switching to the mid-range AMD Wraith Spire cooler, which really is designed for 90 W CPUs only.
In this scenario there's bigger performance differences compared to the AIO. Even with the Wraith Spire running at 100%, there's a significant performance loss in the highly multi-threaded workloads Blender and Cinebench. On the other hand, we're talking about 12% lower performance, which is actually much less than what I would have expected before writing this article. Overall, 2.5% performance loss at 100% fan speed isn't "big", especially games show only tiny differences. Also these 2.5% at 100% line up nicely with the 2.5% that we saw on Noctua 20% fan-speed, so it looks like testing these two coolers covers the whole range of cooler capabilities.
Once we go to lower RPMs, the application performance differences get bigger, especially the Wraith Spire at 40% is noticeably slower. Or is it? 25% in certain apps is probably "noticeable". On the other hand, 7% in Photoshop really makes no difference.
In gaming, we can definitely see bigger FPS losses than on the Noctua, but these are still not big enough to worry about. Maybe Far Cry 6 at 20% fan speed, at 1440p and below, which loses 10% perf—still not the end of the world. All the other games are fine.
Power Draw
Some of our readers were curious about the power draw with these coolers at the various fan settings, so I added the charts below.
As you can see, especially in multi-threaded, the CPU is able to regulate its own power draw greatly. Down to 100 W from 225 W, with a relatively small loss in performance is quite impressive. What's happening here is that due to the lower operating frequency, the power used goes down (not the other way round). Also, due to the way the voltage-frequency curve works on all modern processors, when running at lower frequency, a lower voltage can be used, too, which further reduces power draw.
Conclusion
It's been a fascinating experience trying out the Ryzen 9 7950X with three very different kinds of cooling solutions. The 95°C load temperature of Zen 4 desktop processors is somewhat misunderstood due to the way it's being referred to on social media and online forums. This isn't a T-junction max temperature in the classical sense. Your CPU won't get damaged, or the PC won't turn off the moment it hits this temperature, but rather the processor will aim to run at this temperature, which is the highest safe operational temperature, and adjust its boost frequency and voltages accordingly, to keep the processor at this temperature (not under this temperature). If your cooler is good enough that the processor can boost up to its top 5.75 GHz highest boost frequency, then by all means you'll get this speed. If not, you'll get a lower boost frequency. How much lower? Depends on the cooler. Even with the weakest configuration in our test (Wraith Spire at 20%), you still get 5.3 GHz, which is 92% of the maximum boost clocks. Overclockers can override this limit, and let the processor heat all the way up to 105°C, at which point the processor will unconditionally shut the system down (thermal-trip).
The frenzy around the 95°C load temperature of the processor had gotten so bad that some predicted that cheap air coolers could "start fires." This clearly won't happen as our testing confirms. The Wraith Spire is a 95 W-capable cooler that probably costs AMD $10-15 to bundle with each PIB package. It's a piece of aluminium with a fan—as basic as stock coolers can get. When paired with the 7950X, the cooler is able to keep the processor away from damage or overheating and runs 100% stable all day. The processor will run at 95°C both when gaming and with multi-threaded productivity load, although the frequency yielded varies with the fan setting. You will get frequencies as high as 5.65 GHz under a single-threaded workload with 100% fan (best case), or as low as 3.22 GHz with a 32-threaded workload at 20% fan (the absolute worst case for this processor short of running it without a cooler). You will lose around 7% performance when averaged across all tests, but when you look at individual tests, you'll see that the performance loss can reach almost 25%. Much less performance is lost in gaming workloads (as much as 10% in CPU-limited games, almost nothing in GPU-limited ones).
Things get interesting with the Noctua NH-U14S, which is among the best single-tower fin-stack coolers out there. With its stock NF-A15 140 mm fan in place, the NH-U14S has a manufacturer-rated cooling capacity of 165 W, but let's round it to exactly match the 170 W TDP of the 7950X. In single-threaded workloads, the NH-U14S is almost able to match the most powerful cooler in our comparison, a 420 mm AIO with a Zen MCM-optimized cold-plate offset. The processor exceeds the 5.70 GHz max boost frequency advertised for this processor, although even with a 100% fan setting, the processor isn't able to keep the processor away from the 95°C limit at load. Interestingly, the temperatures don't exceed 80 °C in gaming workloads with 100% fan; and never touch the 90°C-mark until 40% fan—it only does so when its fan is capped at 20%.
At highly-parallelized 16-32 thread workloads, the NH-U14S is able to keep frequencies well above the 5 GHz-mark until 40% fan. This translates into a 1-2% performance loss averaged across all tests between 40-100% fan settings, and an almost 3% loss at the lowest fan-speed. Again, the most highly parallelized tasks such as Cinebench nT see performance losses nearing 5%. When comparing the 420 mm AIO with the NH-U14S, the former does achieve (negligible) performance gains in all workloads, including gaming, particularly at the resolutions people typically game at. Even in games that are CPU-limited, the differences are well contained to just a few percent.
The Arctic Liquid Freezer II 420 mm AIO CLC is the best pre-assembled cooling solution we could possibly find for the 7950X, before entering DIY liquid cooling territory. The cooler not only has a massive 420 mm x 140 mm (3x 140 mm) radiator, but also a water-block mounting system that optionally can be used in a way that's aware of the way the hottest components on the processor—the CPU chiplets are physically arranged in relation to the IHS, such that the center of the block's heat-absorbing cold plate is directly over the chiplets. The cold plate makes contact with the rest of the IHS, too, including the I/O die. Across all four synthetic workload-classes (1-thread, 8-thread, 16-thread, and 32-thread); the AIO is eking out the highest possible frequencies, including 5.50 GHz at 8-thread (typical gaming), 16-thread (typical content creation), and when all 32 threads are maxed out (think rendering or video encoding), the frequency is held as high as 5.25 GHz—impressive for a 32-thread processor to hold across all 16 cores. This doesn't mean you must have an AIO for the 7950X. The difference is pretty much negligible with all workloads, when compared to high-end air cooling. The only notable difference is with Blender and Cinebench nT: 1-2%.
In conclusion, AMD put out its recommendation of at least a 240 mm AIO liquid cooler rather conservatively. Any air cooler will provide enough cooling for the Ryzen 9 7950X (and all weaker Zen 4 CPUs) to operate safely and stable, without any damage. AMD themselves guarantee "the processor is designed to run at TJMax [95°C] 24/7 without risk of damage or deterioration." Due to the way Zen 4 processors are designed to run, they will never exceed those 95°C at stock. Instead of overheating or shutting down, the CPU will regulate its clock speeds automagically to stay as close as possible to 95°C.
Our testing in this article shows that the performance losses are minimal, even when pairing the Ryzen 9 7950X flagship with an entry-level cooler that's running at slow fan speed settings. Whether it makes sense to pair a $700 CPU with a $10 cooler is another question, but it opens up interesting possibilities. When you order a new Zen 4 CPU + watercooling, you don't have to wait for the cooler to arrive. You can simply get started with whatever AM4 heatsink you have lying around and you'll be safe and almost get maximum performance. AMD was wise to make their new Socket AM5 compatible with AM4 coolers, so you have plenty of choice, also on the used market.
The biggest problem is probably psychological. For years we have been trained that "95°C is bad". This is no longer true. 95°C is the new 65°C. The fact that the CPU will always run at around 95°C will make it difficult to quantify a cooler's capability though. Imagine—you're spending $200 on an AIO and you're going from 95°C to 95°C—you'd be disappointed. What you might not be noticing immediately is that with Zen 4 you will now get higher performance in return, automatically, without doing anything, that will be your benefit of upgrading—more perf, not lower temps. As we've seen in this article, if you already have very decent cooling, then the gains from going water are relatively small, and probably not worth it, especially in games. Also, infinitely good cooling will not scale clocks and performance up automatically. Rather there are some hard limits, above which the CPU will not boost, even with excellent cooling.
Overall this is great news for Zen 4 buyers and owners: just stop worrying about the heatsink, buy anything reasonably decent, it doesn't have to be an AIO, and you'll be fine. I do feel that limiting the processor's power draw through Eco Mode or adjusting the PBO limits just to run at lower temperatures isn't worth it. You are sacrificing performance for the peace of mind that you're operating below the temperature that AMD guarantees is the optimum temperature. Limiting power to conserve energy is another thing of course and perfectly reasonable. Early results indicate that a lot of efficiency can be gained from this approach—more on this soon,