The Ryzen 5000 series CPUs from AMD landed a few weeks ago, and after I’ve had about two dozen liquid nitrogen (LN2) overclocking sessions with chips with various core counts, the records are starting to fall. These chips are just flying!
I would go as far as to say the marketing hype and multi-day-info-trickle that was the release actually undersold how good these processors are. I'm not just comparing Team Red vs Team Blue here, but Team Red mainstream vs Team Red high end desktop (HEDT) processors. In fact, the Ryzen 9 5950x on liquid nitrogen competes with water-cooled 64-core Threadrippers in the GPUPI benchmark for CPU 1B! This is jaw dropping and helps make it one of the best CPUs you can buy. More on this later. Let’s see how these Ryzen 9 5950x’s perform from an extreme overclocker’s perspective.
|CPU||AMD Ryzen 9 5950X 16 Core|
|Motherboard||ASRock X570 Taichi|
|RAM||G.SKILL 4000C16 RIPJaws|
|CPU Cooler||Enermax Aquafusion 360 AIO|
|Power Supply||Enermax Maxtytan 1250W PSU|
|Thermal Paste||Thermal Grizzly Kryonaut Extreme TIM|
I used the Enermax Aquafusion 360 for binning the chips. It kept the Ryzen 9 5950X in the 80C range at 1.35V and a 4.8 GHz clock rate while still managing to maintain low fan speeds. It handled the 5950X’s 16 cores, which hit around 300 watts under load, without a blink of an eye. As you would imagine, the lower core-count Ryzen models are also not a problem. The Aquafusion’s three fans definitely help to maintain better temps over a long period of time, and they don't allow the loop to get saturated with heat. Also, my son loves the RGB.
I was lucky enough to test half a dozen Ryzen 9 5950X’s, and they all fell within a 100 MHz range of 4.75 to 4.85 GHz for Cinebench R15 on the Enermax water cooler. That’s with a fixed-ratio overclock with no power savings enabled.
I used the new Thermal Grizzly Kryonaut Extreme thermal interface material (TIM) for all of my testing, and will continue to use it for the time being. The company uses extra machine grinding to make the metallic particles smaller and more uniform. That provides the thinnest possible spread while still maintaining a crazy W/mK value. It’s expensive stuff, but, naturally, the extra processing time leads to higher production costs.
For Vermeer, aka Zen 3 Ryzen 5000, we now have a grouping of eight cores per CCX, instead of a four-core grouping like the previous-gen Matisse chips. The Ryzen 9 5950X has two CCDs that each contain one eight-core CCX. In contrast, a previous-gen Ryzen 9 3950X has two CDDs with two CCXes apiece, each containing four cores.
The new unified eight-core CCX alone is an efficiency boost because those threads all need efficient communication with the L3 cache, and each other. So now you don't have to worry as much about an application spilling threads across multiple CCXes as much. Since the chips now have six to eight active cores in a CCX, there isn’t a need to traverse the Infinity Fabric link quite as much. Now core-to-core communication can occur without going out to the IO die if the chip only uses a single CCD (6/8C CPU), which helps in a myriad of ways.
What does this mean for us overclockers? With only two CCXes to manage instead of four, this also streamlines per-CCX independent overclocking. I have found that CCD 0 will always overclock better than CCD 1, by as much as 125 MHz in some cases. I would assume this is by design, which makes perfect sense.
If you think about PBO and AMD’s other features that boost clock speed, they will use the higher-binned chiplet CCD 0 for high frequency, and if more threads are called, the processor will drop frequency a bit and load the CCD 1 threads while maintaining the power envelope figures. This technique gives you the benefit of very fast single and low core count performance, and the ability to just throw a ton of threads at greedy programs that require over eight.
The benefit to overclocking is I can set, say, CCD 0 to an overclock of 4850 MHz and dial in a 4800 MHz overclock on CCD 1. This increases my Cinebench R15 score by 100 points instead of being limited to the max of CCD 1, which is 4800 MHz. At 4850 MHz on CCD 0 and 4800 MHz on CCD 1, we reach a score of 5470 marks in Cinebench R15. The Enermax Aquafusion 360 keeps the processor around 80C with 1.34V to the core, which is pretty solid.
I really love doing these articles with Tom’s because the large reach helps me shed some insight on the XOC (extreme overclocking) part of things. I have noticed in the comments that people tend to think that LN2 overclocking ruins the processors I test. Shockingly enough, it rarely causes any issues as far as chip degradation goes.
If you think about boost clocks for a minute, if your processor in some instances is hitting up to 1.4-ish volts on auto settings, dialing in 1.65V on LN2 is not really that high, is it? Especially at -192C compared to a normal ambient cooling temperature of, say, 24-28C.
Let’s be clear, if any part I own degrades or gets hurt in any way I am gutted for days about it. This is not a “bench on LN2 then throw away” type of deal. I grind the same CPU over and over and over. I’m chasing records, after all.
For liquid nitrogen overclocking, the Infinity Fabric is still the “limiting” factor. The cold bug (where the processor no longer functions - 00 post code, meaning dead) is very chip-dependent, but in general, a lower fabric clock equals a better ability to use colder temperature. A chip can often run -192C full-pot (meaning the LN2 pot is full, providing maximum cooling) at a 1400 MHz fabric, but then have a cold bug at -125C with a 1600 MHz fabric.
So we have to bin the CPU not for core frequency, but for its ability to run as cold as possible with the fabric dialed in as high as possible. Out of the five Ryzen 9 5950x’s I have tried, only one is cold-bug-free with the fabric set at 1600 MHz. I think it’s a safe estimate to say that around 20 to 30% of the chips will do full-pot overclocking.
Back to the GPUPI benchmark for CPU. I managed to take the 16-core world record at a whopping 6050 MHz with the memory screaming at DDR4-4700 with 14-14-14-14-1t timings. The fabric limit was a little over 1600 MHz. This required 1.65V to the CPU core, but I only needed 1.3V of SoC voltage. The pot temperature idled at -192C, and under load it heated up to around -188C. These chips are not crazy hot, at all.
What makes this even more impressive is changing the perspective from 16 core scores to include any core count. We can see the Ryzen 9 5950X is better than the best Threadripper 3970X, even with both on liquid nitrogen. The 5950X also beats many Threadripper 3990Xs on custom air and water cooling, and is only beaten by LN2-cooled 64-core Threadripper 3990Xs. The 5950X’s IPC and overclockability, as well as the optimized CCX, all come together to shine bright in terms of performance.
The sheer performance I saw with the Ryzen 9 5950X still has me shocked that AMD was able to pull this off. In a way, the incremental upgrades we have seen for so long have left us eager for bigger jumps, and it’s almost like we are being spoiled by a new line of CPUs this time.
It honestly feels like AMD has skipped a generation and given us the meat and potatoes right away instead of another appetizer; I can’t wait till we can have dessert (Threadripper 5000, anyone?).
If you haven't picked one up, or tried to pre-order from the next batch of stock already, the Ryzen 9 5950X is the real deal and I would highly suggest checking it out if you are in the market for an upgrade (see our article on where and how to buy Ryzen 5000 series), or just want to have some overclocking fun. Happy overclocking!