[ad_1]
Here it is – the limit of the Zen 2 architecture on the AM4 platform: 16 cores, 32 threads, turbo mode up to 4.7 GHz, a total cache of 72 MB, and support for DDR4-3200 memory.
For those who are not yet familiar with all aspects of Zen 2, it is recommended to study the past material, and this review will focus on exploring the potential. Let’s get down to business.
Test methodology and test benches
Platform AM4:
- Processor: AMD Ryzen 9 3950X;
- cooling: CBO circuit, CPU water block EK-Supremacy full Nickel;
- thermal interface: Arctic Cooling MX-4;
motherboard: ASUS ROG Crosshair VIII Hero Wi-Fi (AMD X570, UEFI 1105, AGESA 1.0.0.4); - memory: HyperX Predator HX441C19PB3/8 x2 (DDR4-4266, CL19-26-26-45 2T);
- video card: MSI Radeon R9 Nano;
- drive: Kingston KC2000 250GB (SKC2000M8250G);
- power supply: Rosewill Hercules (1600 W).
Platform LGA 1151v2:
- Processor: Intel Core i9-9900K;
- cooling: CBO circuit;
- thermal interface: Arctic Cooling MX-4;
- motherboard: ASUS ROG Maximus XI Gene (Intel Z390, UEFI 1302);
- memory: HyperX Predator HX441C19PB3/8 x2 (4266MHz, CL 19-26-26-45 2T);
- video card: MSI Radeon R9 Nano;
- drive: Kingston KC2000 250GB (SKC2000M8250G);
- power supply: Rosewill Hercules (1600 W).
In liquid nitrogen overclocking mode, NVMe drives worked through an adapter in the PCI-E slot, functioning from the chipset, a discrete card with USB ports for connecting peripherals was used, and an energy-efficient Nvidia GeForce GT 1030 graphics adapter was used. Liquid nitrogen vaporizer: Kingpin F1.
The CBO circuit consisted of a HAILEA HC-300A chiller, a 15-liter liquid tank, two Hydor Seltz 1200 pumps and a CPU EK-Supremacy full Nickel water block.
Software used: Microsoft Windows 10 LTSC (1809) x64, AMD Radeon Adrenalin Driver 19.10.1, AMD Chipset Software Driver 1.11.22.454, Intel Chipset Software 10.1.18, MSI Afterburner 4.6.0, RTSS 7.2.1, HWMonitor 1.41.
Benchmarks were passed at default settings, as well as during overclocking. Games were only tested at 1280×720 resolution on low quality settings. Each test was run five times, the results are the arithmetic mean, the same goes for games with built-in benchmarks. In games without a built-in test, telemetry was recorded using the Riva Tuner Statistics Server. In CS:GO – a standard benchmark on the FPS Benchmark workshop map, in DOTA 2 – a replay of the final game OG vs Liquid at The International 2019, the interval of 19:45–21:15 on an automatic camera, you can see how it looks here. In tests using liquid nitrogen, the strongest of the results obtained was taken into account. Temperature monitoring was carried out using CPUID HWMonitor 1.41, as well as a Benetech GM1312 thermometer with K-type thermocouples.
RAM
For some reason, few people pay attention to this question, but the overall efficiency of the platform is sensitive to RAM. At 16 cores, the influence takes on a tangible effect. The issue is solved either by physically high-speed memory of the DDR4-3600 level and higher, or by a slower memory reserve (at least 32 GB). Here is a quick demo of the situation using Geekbench 5 as an example:
Two quartets of Samsung B-die chips will help us to dot the “i”: 4×8 GB HyperX Predator HX442C19PB3K2/16, as well as 4×16 GB G.Skill Trident Z RGB F4-3200C16D-32GTZR. DDR4-3200 CL16-18-18-38 1T modes with different sizes and number of dies will be tested, as well as high-speed DDR4-3800 CL16-17-16-32 1T + sub-timings with two 8 GB dies each.
Single-core loads are not affected by this issue.
The renderer is already starting to show this problem, but it’s still barely noticeable.
And the following benchmarks show the gravity of the situation.
Practice has confirmed the existence of two scenarios for dealing with this feature. Either the user focuses on the competent setting of high-speed memory, not attaching importance to the volume, or follows the path of increasing the volume without bothering with the settings (for example, using the XMP profile).
Overclocking on liquid nitrogen
Recently, something no one considers it necessary to squeeze all the juice out of processors, spitting on common sense. We need to bring this tradition back! We isolate the board and put the liquid nitrogen evaporator.
First, a reminder of the meaning of two concepts:
- Coldbug (CB) – loss of stability of the platform, or its complete freezing / emergency shutdown at a certain low temperature.
- Cold boot bug (CBB) – the impossibility of correctly starting the platform, or loading it at a certain low temperature.
The first thing to note is the features of FCLK operation at negative temperatures. At -40°C and below, FCLK catches CBB (fails to pass POST). At the same time, after the completion of initialization, let’s say, at higher temperatures, you can then calmly continue to work, and drop the temperatures even lower. However, the FCLK frequency range is noticeably limited, it is possible to maintain stability within 1400-1500 MHz, and these are noticeable drops in performance, especially if you have a successful specimen that can master FCLK 1900 MHz.
As with previous generations of Ryzen, sub-zero temperatures help make great strides with RAM. Or simply compensate for the weak potential of the memory controller. Our test subject managed to compress the primary timings by another 3, and also reached the DDR4-5200 mode. As for the frequency, all cores were overclocked to 5.4 GHz. Scaling relative to voltage is not significant, only 1.55 V per processor and 1.4 V per SoC were enough to conquer these brackets. In comparison, the same Ryzen 5 3600 required 1.85V per processor and 1.5V per SoC. The main limiting factor is temperature. By the way, about her – unlike the younger representatives of the line, like the Ryzen 5 3600, our copy holds the CB line at -150 … -160 degrees. This is a very harmful boundary, in this range a phase transition occurs, when liquid nitrogen evaporates very sharply, and the glass takes on a temperature of -195 ° C. Therefore, during the session, the temperature was kept in the range of -140 … -150 degrees.
There is also a problem with the stability of the PCI Express processor lines, approaching the -100 ° C region, video cards and NVMe drives begin, in a figurative sense, to “fall off”. Fortunately, this is easily solved by transferring everything necessary to the chipset’s PCIe lanes. Being at a decent distance from a glass of liquid nitrogen, they continue to work at positive temperatures.
The comparison will be with the Intel Core i9-9900K, an instance of the author, capable of maintaining the stability of all cores at a frequency of 6.2 GHz, and the cache at 6.1 GHz.
As a counterbalance, there will be a “civil overclocking” Ryzen 9 3950X using CBO at a frequency of 4.5 GHz, FCLK was 1900 MHz, the memory functioned at 3800 MHz with timings of 16-17-16-32-1T.
It is clearly seen how significant the effect of the lack of FCLK frequency at high frequencies is. In tests with full use of all cores, liquid nitrogen does not really help to get ahead in mixed loads, especially those where memory and cache are actively used. Therefore, do not worry if your processor cannot conquer the conditional 4.4 GHz with all active cores.
Synthetics
AMD positions this processor as a top desktop processor, so the comparison will be made with the Intel Core i9 9900k. HEDT platforms are not involved in the tests.
As a result, we see a significant advantage in favor of the Ryzen 9 3950X.
Voltage and frequency dependencies
Surely a good half of users will not squeeze the last juices out of this processor, but will optimize it. As testing with liquid nitrogen showed, a high processor clock speed does not always guarantee a linear increase in performance. Therefore, we will look for the minimum voltage for stable operation 24/7.
In all cases, the SoC voltage was 1.1 V. Stability was tested with a mixture of benchmarks HWBOT x265 – 4k, LinX 0.7.0 AMD (Problem size 37712), as well as xmrig 5.1.1 miner.
Our copy can reach 4.1 GHz with a voltage of only 1.15 V. This allows us to keep temperatures up to 70 degrees under the AMD Wraith Prism cooler. And this is on automatic fan speed control on all 16 cores in a stress test! In typical use, this temperature is unlikely to get close to this milestone, unless the user decides to install such a processor in a Mini-ITX case. By the way, about compactness and power …
16 cores in Mini-ITX
It would seem that this is impossible, but no, it is achievable, and without explosions and overheating. The MSI B450 Gaming Plus AC motherboard was taken for testing. Separately, I would like to note that the AMD B450 system logic set easily mastered the identical memory settings and FCLK frequencies for the X570 chipset.
Attention! The user will have to optimize for a specific instance of the processor, because the board sets as much as 1.472 V (!) per processor by default. Of course, nothing terrible will happen, the processor simply will not use the turbo mode, being at the factory frequency of 3.5 GHz. But why make this compromise, if by spending a couple of hours on optimization you can avoid a performance penalty in the further conditional 5 years of use?
An example of such an optimization: the processor voltage is 1.15 V, the SoC is 1.1 V, the RAM operates at 1.45 V. The processor frequency: 4.0 GHz for all cores, FCLK was 1900 MHz, the memory operates in DDR4 mode -3800 with CL16-17-16-32 1T. The platform was sent to mining for almost 9 hours, as a result, the peak temperature of the processor was 70 degrees, the VRM of the motherboard was 76.
Do not consider compact builds on the Ryzen 9 3950X as stupid by default. This is a viable option for operating the platform with a competent approach to implementation.
Mining
Now in more detail about mining, because a reasonable question probably arises, why did such tests appear in the article? At the end of November, the Monero (XMR) cryptocurrency updated the algorithm on RandomX, in which the Zen 2 architecture performs better than all solutions on the market, leaving behind even previous generations of server platforms with multiple processors. I could not resist, and decided to use the xmrig 5.1.1 miner as a benchmark, because it combines many nodes (RAM, cache, processor) to the maximum load for their work. The issues of the payback period and daily profit will not be considered, because. The cryptocurrency market is dynamic, and we are interested in performance.
Practice shows that the frequency of the CPU can be neglected in favor of lowering the voltage on the processor, reducing heat and consumption.
Games
Ryzen 9 3950X will be tested in three modes:
- factory, DDR4-3200 CL16-18-18-38 1T memory, FCLK = 1600;
- factory with SMT disabled, DDR4-3200 CL16-18-18-38 1T memory, FCLK = 1600;
- overclocking, processor frequency 4.5 GHz, DDR4-3800 CL16-17-16-32 1T memory + sub-timings, FCLK = 1900.
Quite unexpectedly, some games perform worse in overclocked mode than on automatic settings, in particular Gears 5.
As you can see from the results, there is still a difference, eSports fans will be able to feel it on high-frequency monitors, and those who like to play for fun should choose a processor that suits their budget.
conclusions
So the time has come when the performance of simple systems has caught up with the HEDT class, and even surpassed it. Just recently, in 2017, four cores were the popular choice, and eight in the HEDT class. And the process of increasing the number of cores has clearly not yet reached the limit.
Not without drawbacks, too. The main problem is a serious deficit for this model, which is still not over. The time has not yet come for perfect automation out of the box, this processor is best bought only by advanced users, as well as enthusiasts. Without devoting time and effort to individual tuning, in the worst case, you can lose up to 30-35% of potential performance! Forcing the FCLK = ½ MEMCLK mode when operating at a frequency above 1900 MHz also does not allow to achieve more, although over time the restriction may be removed. With games, the competitor still has the lead, although the gap has become smaller, and in some projects even parity.
If the system is assembled as a universal one, for a range of tasks with a margin of safety for the future, the AMD Ryzen 9 3950X is an uncompromising solution. The performance is very high, while the system does not turn into a kind of voracious heating radiator, as AMD FX processors used to be. Do not forget about physical security, Zen 2 does not need any patches and patches that cut the overall performance.
Now it’s up to Intel, look forward to hitting back with the LGA 1200!
Thanks to the Telemart.ua online store for providing the AMD Ryzen 9 3950X processor and the G.Skill memory kit, as well as ASUS Ukraine for the ASUS ROG Crosshair VIII Hero (Wi-Fi) motherboard.
[ad_2]