Zen 2 in all its glory

AMD is moving decisively and quickly, and time after time, it takes noticeable steps towards improving its offerings. And the results are not long in coming. The company is steadily increasing its share of the processor market, and second-generation Ryzen deservedly enjoy a reputation as the best offering for the mass market in terms of price and performance. And even the most notorious skeptics today admit that AMD managed to seriously shake up the processor market and at least make it so that the announcements of new processors turned from ordinary updates into major events in the computer industry.

But now the company wants even more. While Intel continues to torment the 14nm process and Skylake microarchitecture from 2015, AMD is about to finally seize the initiative. The third-generation Ryzen is tasked with demonstrating AMD’s technological superiority and moving it from second place to first place. But will the third attempt by a resurgent AMD to create the best desktop processor in recent history succeed?

We write these lines when we already know the test results. And we can say for sure: the Ryzen 3000 has a lot of positive changes that put it head and shoulders above its predecessors. However, at the same time, problems remain, due to which a substantive story about new products is not too simple.

For this reason, we have divided the material on testing the Ryzen 3000 into two parts. In the first part, we will talk about the new octa-core Ryzen 7 3700X, which is the easiest design to analyze in comparison with the previous generation Ryzen and Intel Core. The second part, which will be released after the first, will be devoted to testing the 12-core flagship Ryzen 9 3900X, with which AMD is going to set a series of absolute records in the mass market segment.

⇡#What you should know about the new Zen 2 microarchitecture

If with the release of the first and second generation Ryzen processors, AMD wanted to convey the idea that it was finally returning to the big leagues of developers and manufacturers of x86 processors, then today’s announcement of the Ryzen 3000 carries with it a completely different message. The company now has a much more ambitious goal of becoming the leader in the processor market that offers the fastest, most energy-efficient and most technologically advanced chips.

And this task does not seem impossible. Over the past year, AMD has managed to build a very solid foundation from which it is quite capable of confidently launching into the sky. Through collaboration with Taiwan’s TSMC, one of Taiwan’s leading contract semiconductor manufacturers, the company is the first in the PC industry to convert its processors to 7nm technology, allowing it to increase die density, raise their operating frequencies, and improve power efficiency at the same time. In addition to this, AMD has introduced another innovation and switched to a new multi-chip (chip) processor layout, which involves assembling end products from several semiconductor crystals, which allows to bypass many manufacturing difficulties and significantly reduce the cost of complex multi-core processors.

But third-generation Ryzen is not only aiming so high because they are able to offer users a lot of high-frequency cores for relatively little money. Something similar has already been in the assortment of AMD before. But there were many complaints about the company’s past processors related to low single-threaded performance, serious delays in inter-core interaction, and an inefficient memory controller. Now all these shortcomings must be eliminated to one degree or another. When it comes to the performance improvement of the new Ryzens over its predecessors, AMD is operating in double digit percentages, and that really does seem like a very big step up from how Intel processors have evolved over the past years.

However, you need to understand that such a significant increase in performance in 3rd generation Ryzen is largely due to the low base effect. The microarchitecture of the new processors is not something fundamentally new: Zen 2 differs from Zen / Zen + only in details and actually brings with it a set of fixes for the most critical problems of its predecessors. But since there were many problems of various kinds, and many of them caused quite significant damage to the overall efficiency of the microarchitecture, their elimination ultimately leads to a noticeable increase in performance.

And yet, we would not want to belittle the merits of AMD. As a result, many positive changes took place in Zen 2: the throughput of all the main intra-processor highways increased, the load on the computing resources available in the processor cores increased, the amount of data with which the processor can operate locally became larger, and the key indicator of the specific performance of the microarchitecture increased significantly — the number of instructions executed per clock (IPC).

We devoted a separate material to a detailed analysis of architectural innovations and improvements in the Ryzen 3000, which describes the structure of the Zen 2 microarchitecture in great detail.

Here we only recall the main reasons for the notorious growth of the IPC indicator. You should know about them at least in order to better understand the test results of representatives of the Ryzen 3000 family. So, these are:

• Increasing the width of the floating point unit (FPU) from 128 to 256 bits. Thanks to this, Zen 2 can execute 256-bit AVX2 instructions in one go, that is, twice as fast as before.
• A twofold increase in the size of the cache of decoded micro-operations, which should reduce downtime of the execution part of the pipeline due to the lack of performance of the x86 instruction decoder.
• Significantly improved branch prediction, the mechanism of which now uses a new TAGE predictor (Tagged geometric) and increased buffers of first and second level branch targets. All this together reduces the probability of branch prediction errors and minimizes the number of situations when the processor is forced to reset the pipeline state due to incorrect code branch predictions.
• The appearance of an additional (third) address generation unit (AGU), which allows executive devices to more timely access the necessary data even at high loads.
• Doubled the width of the cache bus, which also eliminates bottlenecks when accessing data from execution units.
• Doubled L3 cache, totaling 32MB per octa-core chiplet.
• Improved data prefetching algorithms that allow data to be moved from memory to cache before it is requested by code execution.
• Increased scheduler queue sizes to improve the efficiency of SMT technology.
• The increased size of the register file, which gives the processor the ability to process more instructions in parallel without any delay.
• Additional microarchitecture fixes to counter attacks like Specter V4 without compromising performance.

Illustrating microarchitectural improvements with practical examples is easy enough. To do this, we usually use simple synthetic benchmarks from the AIDA64 test utility: they allow us to see how performance has changed when executing certain typical algorithms. In the charts below, we compared the past generation of Ryzen (Pinnacle Ridge) with the current one (Mattisse) using eight-core and sixteen-thread chips running at the same 4.0 GHz clock speed. In addition, the results of the eight-core Coffee Lake Refresh, also running at 4.0 GHz, are placed on the charts.

In fact, all these results are quite interesting. First, they show that in some algorithms the Zen 2 microarchitecture provides almost a twofold increase in performance, while in other cases the performance remains at the same level. Secondly, they allow us to say that from the point of view of relatively simple computational algorithms that are well parallelized and do not need to actively work with external data from RAM, the Zen 2 microarchitecture has not only grown to the efficiency of the Intel Skylake microarchitecture, but even surpassed her.

Matisse shows the most impressive progress in those algorithms that use floating point operations. And more specifically, where the AVX2, FMA3 and FMA4 instructions are applied. After all, it was their execution in Zen 2 that was doubled.

As for integer calculations, there were no particular problems with them in past Ryzen processors either. Now there has been only a slight change in performance, primarily due to changes in caching and decoding of instructions: with a decrease in the volume of the L1I cache and with an increase in the cache of decoded micro-ops. Separately, the relatively low result of Matisse in the Photoworxx CPU test should be noted. The fact is that this is the only benchmark in which, among other things, the performance of the memory subsystem plays a role. And with it, the new Ryzen is really not as good as with the microarchitecture. However, let’s not get ahead of ourselves.

⇡ # Infinity Fabric bus and inter-core communication speed

If we talk about eight-core and six-core processors, the third-generation Ryzen processors have retained their traditional basic structure — they are composed of two four-core CCX (Core Complex) complexes, which are placed in one eight-core CCD (Core Complex Die) processor chip and connected inside it by a bus Infinity Fabric. However, the difference from past processors is that the eight-core processor is no longer a single monolithic die. The memory controller, PCI Express controller and SoC elements are removed from the CCD chiplet and assembled into a separate I / O chiplet manufactured using 12nm technology at GlobalFoundries enterprises. At the same time, such a two-chip layout does not affect the connection between the cores and the L3 cache in any way — everything remains the same here.

In processors with 12 and 16 cores, one more level of hierarchy is still added — they use similar eight-core CCD chipsets, but in double quantity. At the same time, CCD chipsets do not have a direct connection with each other. They are connected by the Infinity Fabric bus only to the I/O chiplet, so all interaction between the cores located in different chiplets occurs through an intermediate node — the I/O chiplet.

In the end, it turns out that even in the case of eight-core processors, the cores are unequal in relation to each other: there are “close” cores (located in the same CCX complex), and there are “distant” cores (located in different CCX and able to communicate with each other). other only through Infinity Fabric). In processors based on a pair of CCDs, there are also «very distant» cores — physically located in different crystals. Due to this specificity, the delays in the exchange of data between the cores are different depending on whether they are combined in one CCX or located in different ones. And this is a rather alarming moment: in past generations of Ryzen processors, the delays that occurred when cores from different CCXs communicated became quite noticeable and in some cases slowed down performance.

In Ryzen 3000, this problem should have been partially fixed. First, AMD worked with Microsoft and was able to ensure that the operating system scheduler now takes into account the processor topology and first of all loads the cores from one CCX complex, moving on to the next CCX only when the free cores in the previous one are already loaded with work. This strategy is inherent in the scheduler in the new version of Windows 10 May 2019 Update, and in relation to Ryzen processors, this allows you to reduce the number of inter-core accesses on the high-latency Infinity Fabric bus and concentrate calculations, if they do not load all processor cores, inside the smallest processor structural unit.

Secondly, the Infinity Fabric bus in the new generation Ryzen is noticeably accelerated in itself: its width has doubled — from 256 to 512 bits. Does it improve the situation much? The positive effect is easy to verify, for this we performed our traditional test of delays that occur when transferring data between cores. For comparison, below are the results of measurements made not only for the third generation Ryzen octa-core processor, but also for the previous generation octa-core processor (Pinnacle Ridge), as well as for the Coffee Lake Refresh octa-core processor. All processors during the test were brought to a single clock frequency of 4.0 GHz, the memory of all CPUs worked in DDR4-3466 mode, which means that the Infinity Fabric bus in the compared Ryzen used the same frequency of 1733 MHz.

The situation in the Ryzen 3000 has indeed improved markedly. Cores belonging to the same CCX complex are now able to communicate 25% faster, and cores belonging to different CCXs are one-third closer to each other. Thus, Ryzen 3000, at least if we talk about processors with no more than eight cores, will be much less susceptible to problems with high latency during inter-core interaction. Moreover, in terms of the speed of communications between cores belonging to the same CCX complex, the new members of the Ryzen family surpassed even Coffee Lake Refresh, which uses a ring bus, which is considered the most successful way to connect processor components into a single whole.

The positive impact of the high speed of Infinity Fabric should be manifested not only when transferring data between cores. It is worth recalling that each CCX complex in Ryzen processors has its own L3 cache, and the massive 32 MB L3 cache in the eight-core Ryzen 3000 is actually two caches up to 16 MB. Therefore, calls through Infinity Fabric also occur when the cores of one CCX complex need data located in the L3 cache of the second CCX complex. Therefore, the observed acceleration of Infinity Fabric should have a positive impact on performance in a fairly wide range of situations, including when actively working with data.

However, another problem related to the speed of Infinity Fabric remained unresolved: the frequency of this bus continues to be associated with the frequency of the memory controller. Although AMD has implemented the Infinity Fabric asynchronous mode in the new processors, the frequency of this bus still cannot exceed the frequency at which the memory controller operates, which means that the choice of DDR4 SDRAM modules will continue to have a noticeable impact on the performance of the Ryzen 3000.

⇡#Cache speed

If we talk about working with data, then the cache memory subsystem in the Ryzen 3000 processors has not changed much. The cache memory of the first (L1D) and second levels has retained the same size, organization and latencies, and the only innovation is the increased cache of the third level. Due to the transition to 7-nm technology, AMD allowed itself to significantly increase the transistor budget of CCX-complexes, and thanks to this, the third-level cache was doubled — up to 16 MB for every four cores. Even so, the area occupied by the Ryzen 3000 CCX complex on a 7nm semiconductor chip is only 31.3mm.²while in the previous generation processors, which are manufactured using 12nm technology, the CCX complex occupies 60mm².

But the increase in the volume of the L3 cache was not so much due to the generosity of the developers. This is partly a forced measure. In new processors with a chiplet layout, the memory controller has moved away from the computing cores, and caching the largest possible amount of data is a technique that is needed in order to try to reduce the number of situations when processor cores are idle waiting for data to be received from memory. AMD representatives say that first of all this should help solve performance problems in games, but we’ll check it out.

Now I want to talk about another point: the growth of cache memory is always accompanied by an increase in its delays. This is what happened this time, but in fairness it should be noted that the increase in latency turned out to be quite small — from 38-39 to 41-42 cycles.

In the graphs below, we compared the cache latency of 2nd and 3rd generation Ryzen eight-core processors, as well as current representatives of the Intel Core family. All processors during the measurements were brought to a single frequency of 4.0 GHz.

The L1 and L2 caches in the Ryzen 3000 have not changed their key performance parameters compared to the previous generation processors. The latency of L1- and L2-cache remained at the level of 4 and 12 cycles. However, it would be wrong to say that the cache memory closest to the computing cores has not changed at all. The Ryzen 3000’s L1 cache is actually faster, as it is now capable of serving two 256-bit reads and one 256-bit write per clock cycle, which means twice the throughput of previous Ryzen family processors.

As a result, the speeds of the L1 and L2 caches in the Ryzen 3000 have become fully comparable with the speeds of the lower-level cache memory in the competitor’s current mainstream processors. And the L3 cache in the new Ryzen, although it has increased its latency, can still offer lower latency compared to the L3 cache in Intel Coffee Lake Refresh processors. However, one should not lose sight of the fundamentally different algorithms for the operation of the L3 cache in processors from different manufacturers. In Zen/Zen+ and Zen 2, the L3 cache is very simple and victimized, and also independent for each CCX complex. At the same time, Intel’s consumer processors for the LGA1151 platform implement a smarter, inclusive write-back cache that is shared across all processor cores. In other words, the practical efficiency of the L3 cache in AMD and Intel processors is very different.

At the same time, the given latency graphs give certain reasons for concern. Namely, the final part of the latency curve for Matisse, which shows the characteristics of the memory subsystem, causes concern. As you can see, there are no reasons for optimism here: the third-generation Ryzen turned out to be worse than their predecessors and, as a result, they seriously lose to competitor processors in terms of memory latency. What’s the matter?

⇡ # Working with memory

The chiplet layout implemented in the Ryzen 3000 separated the processing cores and the processor memory controller. While CCX complexes with cores and L3 cache are located in 7nm CCD chipsets, the memory controller, together with the PCI Express controller and SoC elements, is placed in another I / O chiplet. The connection between the chiplets mounted in the processor on a single textolite substrate occurs using the Infinity Fabric bus, which means that an additional stage has appeared on the way of data from memory to processor cores. And although AMD says that the external Infinity Fabric bus is similar in speed to the bus connecting CCX complexes inside the CCD chiplet, all this, one way or another, should have affected the delays that occur when accessing memory.

In other words, when it was discovered that the memory latency in the Ryzen 3000 was worse than before, we were not at all surprised. Another thing is more interesting: how much the speed of working with memory has worsened in the new AMD processors. The results of the Cachemem test from the AIDA64 utility clearly answer this question (for correct measurements, all processors are set to a single frequency of 4.0 GHz; in all cases, dual-channel DDR4-3466 SDRAM with timings of 16-16-16-36-1T is installed in the systems).

As can be seen from the above data, the memory latency deterioration in the Ryzen 3000 compared to the previous generation processors is about 11%. In addition, throughput indicators have also deteriorated: the write speed shown by the Ryzen 3000 memory controller has become one and a half times lower than it was before. In other words, miracles do not happen: just as the shift of the memory controller from the chipset to the processor in the mid-2000s accelerated memory processing, the reverse separation of the memory controller from the computing cores naturally led to the opposite result.

Moreover, for AMD processors, the increase in memory latency is not just an annoying trifle, it is really a very unpleasant moment. In terms of the speed of working with memory, Ryzen of the previous generation already noticeably lost to competitor processors. Now, with the release of Ryzen 3000, the situation is only getting worse. Although the read and copy throughputs of Ryzen 3000 and Intel Coffee Lake Refresh remain comparable, in terms of write speed and memory latency, the new AMD processors are 1.6-1.8 times inferior to competitors.

But not everything is so dramatic. There is some good news for potential third-generation Ryzen buyers. Most importantly, the new processors use a significantly redesigned memory controller, which is far from being as whimsical as its predecessor. This is also reflected in the passport characteristics: the new Ryzen 3000 received official support for DDR4-3200 SDRAM, which was not formally offered before. Moreover, memory performance in DDR4-3200 mode is guaranteed for any pair of modules, regardless of their organization and base component.

In addition, if we talk about the capabilities implemented in the new memory controller, then it is worth mentioning a couple more important things. First, the Ryzen 3000 will now support 32GB modules, which means a total of 128GB of memory can be installed in systems based on the new processors. Secondly, the memory controller supports ECC. However, the ability to use this feature will depend on motherboards, and experience shows that manufacturers usually do not enable it in common consumer platforms.

However, the main advantages of the new controller become apparent in its practical use. Without exaggeration, it can be called problem-free: it is omnivorous in relation to memory modules and is much more stable in operation, without requiring tedious selection of timings to achieve stability at a relatively high frequency. While with Ryzen processors of past generations, memory modules were rarely able to run in modes faster than DDR4-3466, running memory in faster modes does not cause problems with the new controller. Together with the increased volume of the third level cache, this largely compensates for the increase in the latency of the memory subsystem as a whole.

However, AMD would not be itself if a list of limitations and caveats had not been attached to the positive changes. So, despite the possibility of significant memory overclocking, the maximum rational mode of memory operation with Ryzen 3000 is DDR4-3600. It is in this case that maximum performance is achieved, while modes that are faster in frequency are meaningless in terms of performance.

Ryzen 7 3700X with DDR4-3600

The reason lies in the links between the frequencies of the memory itself, the memory controller and the Infinity Fabric bus. They made life difficult for AMD fans in the past and will continue to do the same in Ryzen 3000 processors, although there have been some changes for the better. Most importantly, AMD was able to decouple the frequency of the Infinity Fabric bus from the memory frequency: they can change independently in new processors. However, there is an important nuance: the frequency of Infinity Fabric must be either equal to the memory frequency or less than. This means that the choice of memory modules will continue to have a significant impact on the performance of the processor as a whole.

The second caveat concerns the fact that the maximum allowable frequency of Infinity Fabric in Ryzen 3000 is 1800 MHz, and if you select higher values, the processor cannot function. There is also a third nuance. It concerns the fact that when using memory modules faster than DDR4-3600, the memory controller clock generator automatically switches to 2:1 mode, that is, it starts to operate at half the frequency.

All this in sum leads to the fact that it makes no practical sense to use memory in modes faster than DDR4-3600: when crossing this border, additional and very significant delays are added to the operation of the memory subsystem due to the switching on of asynchrony.

Ryzen 7 3700X with DDR4-3866

As you can see from the screenshot, the memory latency in the DDR4-3866 mode is about 9 ns higher than in the DDR4-3600 mode with the same timing settings. Compensating for such an increase in latency by further increasing the frequency of DDR4 SDRAM, if we talk about the usual, non-extreme overclocking, is almost unrealistic.

There is only a small hope that the frequency of Infinity Fabric in serial processors, under some conditions, can still be raised above 1800 MHz, because in theory, motherboards have an appropriate setting with a wide choice of frequencies for this bus. In this case, in systems based on Ryzen 3000 processors, it may make sense to use modules faster than DDR4-3600.

However, we never managed to cross the 1800 MHz line for Infinity Fabric: choosing higher values ​​inevitably led to the complete inoperability of the test system.

⇡#X570 chipset and compatibility with old boards

We tested the Ryzen 3000 armed with a platform based on the X570 chipset. AMD has prepared this chipset specifically for the release of its processors with the Zen 2 microarchitecture, but the X570 board is a completely optional companion for the new Ryzen. Like their predecessors, the Ryzen 3000 is compatible with the familiar Socket AM4 socket and is capable of running boards released during both the first and second generation of Ryzen.

However, not all so simple. For new processors to work in old motherboards, they need support at the BIOS level, and with its implementation, everything is far from being so smooth for marketing reasons. In fact, Ryzen 3000 will certainly be compatible with any X470 and B450 boards, but with all other platforms, the situation is at the mercy of motherboard manufacturers. Therefore, support for new processors in some specific boards with X370, B350 and A320 chipsets may not appear.

The compatibility criterion is simple: in order for a board to work with Ryzen 3000, its BIOS must be rebuilt using the AMD AGESA Combo_AM4 PI 1.0.0.1 or later libraries. If the motherboard manufacturer has released an appropriate firmware update, the board for the Ryzen 3000 will do.

Nevertheless, it would not be a good idea to use old boards for testing new processors now. The fact is that motherboard manufacturers have thrown all their efforts into optimizing the BIOS of the latest generation of Socket AM4 platforms, and support for Ryzen 3000 in older platforms is implemented according to the residual principle. This is manifested in the fact that almost all available BIOS updates for older boards are based on the AGESA code versions 1.0.0.1 or 1.0.0.2, and these versions do not fully reveal the performance of the Ryzen 3000.

For full-fledged operation of new CPUs and reaching the maximum level of performance in the BIOS code, the AMD AGESA Combo_AM4 PI 1.0.0.3 libraries must be used, and this condition is currently met only for a few boards, mainly with the X570 chipset. This is the reason why we ran our tests with an X570-based board, which, thanks to better optimizations, can offer better performance when paired with Ryzen 3000. However, this situation is temporary: as the BIOS code is updated in older boards, their performance with Ryzen 3000 should be catch up to the same level that fresh platforms provide today.

The X570 chipset itself does not add any features that are especially in demand at the moment to the Socket AM4 platform. The main reason why users should pay attention to it is the appearance of the PCI Express 4.0 interface in boards based on it. If Ryzen 3000 is used in such boards, this interface is supported by both PCIe x16 graphics slots and M.2 slots for NVMe drives, as well as any other PCIe slots. In addition, new generation boards tend to have a large number of USB 3.1 Gen2 ports: the processor and X570 chipset can support up to 12 of these ports.

Ryzen 3000 processors have 24 PCI Express 4.0 lanes. Four of this number are used to connect to the system logic set, four more lines are given to work with the system NVMe SSD. The remaining 16 lines are the interface with the graphics card.

The X570 chipset has 20 PCI Express 4.0 lanes at its disposal, four of which are needed to communicate with the processor. The remaining 16 lanes can be assigned by the motherboard manufacturer to PCIe, M.2 slots or configured as additional SATA ports.

At this stage, all this does not seem to be so in demand, although devices with PCI Express 4.0 support are gradually penetrating the market. So, a promising interface with twice the bandwidth will use the Radeon RX 5700 and RX 5700XT graphics cards. In addition, solid state drives based on the Phison PS5016-E16 controller (for example, Gigabyte AORUS SSD NVMe Gen4 or Corsair Force Series MP600) will also begin to appear on sale in the near future, which will also be able to take advantage of the increased interface bandwidth.

However, if you are considering purchasing an X570-based motherboard, you should keep in mind that this chipset is a very hot chip, with heat dissipation ranging from 11 to 14 watts at peak loads. Technically, it is a reconfigured I / O chip from EPYC Rome server processors, that is, it is based on a 14-nm crystal produced at the facilities of GlobalFoundries. And therefore, it is not at all surprising that, like processors, it needs active cooling: the vast majority of X570 motherboards use a chipset cooler with a fan.

In addition, Socket AM4 boards using the X570 will be among the flagship platforms. And this means that they will cost a lot: we can expect that the prices of the cheapest new-generation boards will start from $200-$250.

⇡#More about Ryzen 7 3700X

The lineup of Ryzen 3000 processors (codename Matisse) consists of six modifications: two six-core, two eight-core, twelve-core and sixteen-core processors. For the first introductory review, we chose the middle model — the junior eight-core processor Ryzen 7 3700X.

This was done primarily because it is easy for him to pick up rivals — both the competitor and Ryzen of past generations have eight-core processors. In addition, the Ryzen 7 3700X seems to be one of the most requested new products. This can be concluded if you look at the composition of the model range in full.

The Ryzen 3000 is attractive due to four factors: high performance, affordable price, moderate power consumption and heat dissipation, and hope for non-zero overclocking potential.

Indeed, the Ryzen 7 3700X is a full-fledged eight-core Socket AM4 processor with Zen 2 microarchitecture, assembled from a single 7nm CCD chip with a full set of active cores and a 12nm I / O chip. It has slightly lower frequencies compared to the older octa-core Ryzen 7 3800X, but the difference in maximum frequency is only 100 MHz. There are no other fundamental differences: the Ryzen 7 3700X has both a full-size L3 cache with a total volume of 32 MB and an L2 cache with a capacity of 512 KB per core.

You can ignore the fact that 3.6 GHz is indicated as the base frequency for this processor — in reality, due to the Precision Boost 2 auto-overclocking technology, the processor almost always reaches a significantly higher speed. For example, when tested in Cinebench R20 with a load of various numbers of cores, our sample showed real frequencies in the range from 4.1 to 4.4 GHz, which not only looks good, but also exceeds the typical operating frequencies of the previous flagship, the Ryzen 7 2700X.

However, at the same time, AMD is clearly cunning when talking about the energy efficiency of the Ryzen 7 3700X and referring it to a 65-watt thermal package. To understand this, just look at how the system behaves with this CPU in nominal mode in stress tests, for example, in Prime95.

Literally everything here raises questions. And you should start with a high operating temperature, which for our copy of the Ryzen 7 3700X in the Prime95 29.8 test reached 90 degrees, despite the fact that AMD itself considers the processor to only heat up to 95 degrees as the maximum possible. But such a picture in our case is observed not even with a boxed Wraith Prism, but with a much more powerful Noctua NH-U14S.

Of course, a processor die made according to 7nm standards has an extremely small “contact patch” with a heat-distributing cover, and, therefore, it is really more difficult to cool Ryzen 3000 than 12- and 14nm processors. However, a 90-degree CPU temperature raises doubts that such a processor under load demonstrates power consumption at the level of 53 W, as reported by all its internal sensors. One gets the impression that AMD deliberately and very much underestimates consumption figures so that the processor automatically overclocks to higher frequencies as part of the technology. precision Boost 2, which for the 65-watt Ryzen 7 3700X sets the upper bar for consumption at 88 watts.

What is the real consumption of the Ryzen 7 3700X can be judged by the sensors of the motherboard power converter. According to their testimony, the processor, which allegedly creates an electrical load of 53 watts, is supplied with a current of 106 watts, plus about 15 watts more to the SoC. The system as a whole at this time demonstrates consumption of the order of 185-190 W, so no doubtdoes not remain: the 65-watt Ryzen 7 3700X under load is capable of consuming about twice the declared thermal package. In other words, the energy efficiency of the Ryzen 7 3700X is a lie, slander and provocation.

Of course, such a situation with consumption can be attributed to the wrong setting of the technology. precision Boost 2 in the BIOS of a particular motherboard or AMD’s intentional neglect of the thermal package, but you need to understand that if the manufacturer decides to return the promised energy efficiency to the Ryzen 7 3700X, then its performance will inevitably suffer. Nothing else is given here.

But what cannot be taken away from the Ryzen 7 3700X is a very attractive price. The Ryzen 7 3700X is not only the cheapest Zen 2 octa-core processor, it is also the lowest cost-per-core processor in the new lineup. In addition, its cost is significantly lower than the price of the competitor’s junior eight-core offer. All of this combined could easily make the Ryzen 7 3700X the “choice of the millions,” regardless of any of its shortcomings.

⇡ # Overclocking

The Ryzen 7 3700X is the youngest octa-core in the lineup of new products, and such introductions usually mean that this processor can be efficiently overclocked, at least reaching the frequencies of the older representatives of the lineup. Moreover, AMD traditionally did not create any obstacles for overclockers. The multipliers in the Ryzen 7 3700X are not blocked, and the heat-dissipating cover continues to be soldered to the surface of semiconductor crystals, despite the fact that there are now two of them under it.

Nevertheless, overclocking is still not at all about the Ryzen 7 3700X. AMD in each new generation systematically squeezed all the juice out of the frequency potential available in the processors and by now has reached perfection in this. We can say that the Ryzen 7 3700X works almost at the limit of its capabilities even in nominal mode due to Precision Boost 2 technology, which is clearly hinted at by the observed operating temperatures.

One way or another, the maximum frequency that we managed to “squeeze” when manually overclocking the Ryzen 7 3700X turned out to be only 4.2 GHz. When the supply voltage was increased to 1.4 V, the processor worked stably at this frequency and passed stress testing in Prime95, however, the temperature under load increased to 105 degrees, which can hardly be considered a normal operating mode.

The result obtained is more of a theoretical value, and there is no point in resorting to such overclocking in practice. The increase in performance under multi-threaded load will be several percent, despite the fact that when the cores are not fully loaded, the processor will work even slower than in the nominal mode.

At the same time, AMD offers enthusiasts another way to increase performance — adjusting the Precision Boost 2 parameters so that the processor independently reaches higher frequencies as part of the built-in auto-overclocking algorithm. This technology allows you to change its key reference constants — current consumption limits (PPT) and electric power limits (TDC and EDC) along with an increase in the upper frequency limit, which can be used for overclocking. However, we did not manage to achieve any noticeable effect by changing these limits in the case of the Ryzen 7 3700X. Even in the nominal mode, Precision Boost 2 manages the frequencies of the Ryzen 7 3700X very aggressively, and the main problem that gets in the way of overclocking is not the restrictions on consumption and currents, but high temperatures.

⇡#Description of test systems and testing methods

To determine the relative performance of the new Ryzen 7 3700X, we compared it to other mainstream octa-core processors on the market. This is the Ryzen 7 2700X, which belongs to the previous generation of processors for the Socket AM4 platform, as well as the Intel Core i7-9700K and Core i9-9900K processors.

Keep in mind that the Ryzen 7 3700X is priced at $329, which makes it a better alternative to the Core i7-9700K, which currently has a $374 MSRP, from a marketing standpoint. However, from the point of view of the same multithreading capabilities, the Ryzen 7 3700X also makes sense to compare with the more expensive Core i9-9900K, which has eight cores and, unlike the Core i7-9700K, supports Hyper-Threading technology, although it costs $488.

In addition, according to available information, prices for Intel processors should be reduced.

In the end, the list of components involved in testing turned out to be as follows:

• Processors:
  • AMD Ryzen 7 3700X (Matisse, 8 cores + SMT, 3.6-4.4 GHz, 32 MB L3);
  • AMD Ryzen 7 2700X (Pinnacle Ridge, 8 cores + SMT, 3.7-4.3 GHz, 16MB L3);
  • Intel Core i9-9900K (Coffee Lake Refresh, 8 cores + HT, 3.6-5.0 GHz, 16MB L3);
  • Intel Core i7-9700K (Coffee Lake Refresh, 8 cores, 3.6-4.9 GHz, 12 MB L3).
• CPU cooler: Noctua NH-U14S.
• Motherboards:
  • ASRock X570 Taichi (Socket AM4, AMD X570);
  • ASRock X470 Taichi (Socket AM4, AMD X470);
  • ASRock Z390 Taichi (LGA1151v2, Intel Z390).
• Memory: 2×8 GB DDR4-3600 SDRAM, 16-16-16-36 (G.Skill Trident Z RGB F4-3600C16D-16GTZR).
• Video Card: NVIDIA GeForce RTX 2080 Ti (TU102, 1350/14000 MHz, 11 GB GDDR6 352-bit).
• Disk subsystem: Samsung 960 PRO 1TB (MZ-V6P1T0BW).
• Power Supply: Thermaltake Toughpower DPS G RGB 1000W Titanium (80 Plus Titanium, 1000W).

All compared processors were tested with memory running in DDR4-3600 mode with XMP timings, with the exception of the Ryzen 7 2700X, which does not work in this mode with the kit we use. It used the slightly slower DDR4-3466 mode.

All compared processors were tested with the settings accepted by the motherboard manufacturers «by default». This means that for both AMD and Intel processors, the established power consumption limits are ignored and the maximum possible frequencies are used in order to obtain maximum performance. It is worth emphasizing that the vast majority of users operate processors in this mode, since enabling restrictions on heat dissipation and power consumption requires special BIOS settings.

Testing was performed on the Microsoft Windows 10 Enterprise (v1903) Build 18362 operating system using the following set of drivers:

• AMD Chipset Driver 1.07.07.0725;
• Intel Chipset Driver 10.1.1.45;
• Intel Management Engine Interface Driver 11.7.0.1017;
• NVIDIA GeForce 430.86 Driver.

Description of the tools used to measure computing performance:

Comprehensive benchmarks:

• BAPCo SYSmark 2018 — testing in Productivity scenarios (office work: spreadsheet processing, archiving and unzipping files, working with PDF and text documents, email, installing and uninstalling programs, creating presentations, optical recognition of a scanned document), Creativity (working on multimedia content — merging panoramas from several images, creating HDR photos, preparing images for printing, importing and exporting photos, face recognition in photos using AI algorithms, video transcoding, preparing videos for publication on the web), Responsiveness (launching «heavy» software packages, working in a browser with a large number of open tabs, installing and removing programs, switching between browser tabs and open applications, writing a set of documents to a folder).
• Futuremark 3DMark Professional Edition 2.8.6546 — testing in the Time Spy Extreme 1.0 scene.

Applications:

• 7-zip 19.00 — archiving speed testing. The time taken by the archiver to compress a directory with various files with a total volume of 3.1 GB is measured. The LZMA2 algorithm and the maximum compression ratio are used.
• Adobe After Effects CC 2019 16.1.1 — Animation video rendering speed test. The time taken by the system to render a pre-prepared video in 1920 × 1080@30fps resolution is measured.
• Adobe Photoshop Lightroom Classic CC 8.2.1 — performance testing for batch processing of a series of images in RAW format. The test scenario includes post-processing and export to JPEG at 1920 × 1080 resolution and maximum quality of two hundred 16-megapixel RAW images taken with a Fujifilm X-T1 digital camera.
• Adobe Premiere Pro CC 2019 13.1 — performance testing for non-linear video editing. It measures the rendering time to YouTube 4K format of a project containing HDV 2160p30 footage with various effects applied.
• Blender 2.79b — testing the speed of the final rendering in one of the popular free packages for creating three-dimensional graphics. The duration of building the final model from Blender Cycles Benchmark rev4 is measured.
• Corona 1.3 — testing the rendering speed using the renderer of the same name. Measures the build speed of the standard BTR scene used to measure performance.
• Microsoft Visual Studio 2017 (15.9.13) — measuring the compilation time of a large MSVC project — a professional package for creating three-dimensional graphics Blender version 2.79b.
• OBS Studio 23.2.1 — testing the performance and smoothness of game streaming. The following video stream settings are used: x264 encoder, resolution 1080p@60fps, bitrate 6 Mbps, CPU Usage Preset = medium.
• Stockfish 10 — testing the speed of the popular chess engine. The speed of enumeration of variants in the position «1q6/1r2k1p1/4pp1p/1P1b1P2/3Q4/7P/4B1P1/2R3K1 w» is measured.
• V-Ray 4.10.03 — testing the performance of a popular rendering system using the standard V-Ray Benchmark Next application.
• VeraCrypt 1.23 — cryptographic performance testing. The benchmark built into the program is used, which uses Kuznyechik-Serpent-Camellia triple encryption.
• x264 r2969 — testing the speed of video transcoding to H.264/AVC format. To evaluate performance, we use the original 2160p@24FPS AVC video file with a bitrate of about 42 Mbps.
• x265 3.1+2 8bpp — testing the speed of video transcoding to the promising H.265/HEVC format. For performance evaluation, the same video file is used as in the x264 encoder transcoding speed test.

Games:

• Assassin’s Creed Odyssey. Resolution 1920 × 1080: Graphics Quality = Ultra High. Resolution 3840 × 2160: Graphics Quality = Ultra High.
• Civilization VI: Gathering Storm. Resolution 1920×1080: DirectX 12, MSAA=4x, Performance Impact=Ultra, Memory Impact=Ultra. Resolution 2560×1440: DirectX 12, MSAA=4x, Performance Impact=Ultra, Memory Impact=Ultra.
• Far Cry 5. Resolution 1920 × 1080: Graphics Quality = Ultra, HD Textures = On, Anti-Aliasing = TAA, Motion Blur = On. Resolution 2560 × 1440: Graphics Quality = Ultra, HD Textures = On, Anti-Aliasing = TAA, Motion Blur = On. Resolution 3840 × 2160: Graphics Quality = Ultra, Anti-Aliasing = Off, Motion Blur = On.
• Grand Theft Auto V. 1920 × 1080 resolution: DirectX Version = DirectX 11, FXAA = Off, MSAA = x4, NVIDIA TXAA = Off, Population Density = Maximum, Population Variety = Maximum, Distance Scaling = Maximum, Texture Quality = Very High, Shader Quality = Very High, Shadow Quality = Very High, Reflection Quality = Ultra, Reflection MSAA = x4, Water Quality = Very High, Particles Quality = Very High, Grass Quality = Ultra, Soft Shadow = Softest, Post FX = Ultra, In -Game Depth Of Field Effects = On, Anisotropic Filtering = x16, Ambient Occlusion = High, Tessellation = Very High, Long Shadows = On, High Resolution Shadows = On, High Detail Streaming While Flying = On, Extended Distance Scaling = Maximum, Extended Shadows Distance = Maximum. Resolution 2560 × 1440: DirectX Version = DirectX 11, FXAA = Off, MSAA = x4, NVIDIA TXAA = Off, Population Density = Maximum, Population Variety = Maximum, Distance Scaling = Maximum, Texture Quality = Very High, Shader Quality = Very High , Shadow Quality = Very High, Reflection Quality = Ultra, Reflection MSAA = x4, Water Quality = Very High, Particles Quality = Very High, Grass Quality = Ultra, Soft Shadow = Softest, Post FX = Ultra, In-Game Depth Of Field Effects = On, Anisotropic Filtering = x16, Ambient Occlusion = High, Tessellation = Very High, Long Shadows = On, High Resolution Shadows = On, High Detail Streaming While Flying = On, Extended Distance Scaling = Maximum, Extended Shadows Distance = Maximum. Resolution 3840 × 2160: DirectX Version = DirectX 11, FXAA = Off, MSAA = Off, NVIDIA TXAA = Off, Population Density = Maximum, Population Variety = Maximum, Distance Scaling = Maximum, Texture Quality = Very High, Shader Quality = Very High , Shadow Quality = Very High, Reflection Quality = Ultra, Reflection MSAA = x4, Water Quality = Very High, Particles Quality = Very High, Grass Quality = Ultra, Soft Shadow = Softest, Post FX = Ultra, In-Game Depth Of Field Effects = On, Anisotropic Filtering = x16, Ambient Occlusion = High, Tessellation = Very High, Long Shadows = On, High Resolution Shadows = On, High Detail Streaming While Flying = On, Extended Distance Scaling = Maximum, Extended Shadows Distance = Maximum.
• Hitman 2. 1920 × 1080 resolution: DirectX 12, Super Sampling = 1.0, Level of Detail = Ultra, Anti-Aliasing = FXAA, Texture Quality = High, Texture Filter = Anisotropic 16x, SSAO = On, Shadow Maps = Ultra, Shadow Resolution = high. Resolution 3840 × 2160: DirectX 12, Super Sampling = 1.0, Level of Detail = Ultra, Anti-Aliasing = FXAA, Texture Quality = High, Texture Filter = Anisotropic 16x, SSAO = On, Shadow Maps = Ultra, Shadow Resolution = High.
• Kingdom Come: Deliverance. Resolution 1920 × 1080: Overall Image Quality = Ultra High. Resolution 3840 × 2160: Overall Image Quality = Ultra High.
• Shadow of the Tomb Raider. Resolution 1920×1080: DirectX12, Preset=Highest, Anti-Aliasing=TAA. Resolution 3840 × 2160: DirectX12, Preset = Highest, Anti-Aliasing = Off.
• The Witcher 3: Wild Hunt. Resolution 1920 × 1080: Graphics Preset = Ultra, Postprocessing Preset = High. Resolution 2560 × 1440: Graphics Preset = Ultra, Postprocessing Preset = High. Resolution 3840 × 2160: Graphics Preset = Ultra, Postprocessing Preset = High.
• Total War: Warhammer II. Resolution 1920 × 1080: DirectX 12, Quality = Ultra. Resolution 3840 × 2160: DirectX 12, Quality = Ultra.
• Watch Dogs 2. Resolution 1920 × 1080: Field of View = 70°, Pixel Density = 1.00, Graphics Quality = Ultra, Extra Details = 100%. Resolution 3840 × 2160: Field of View = 70°, Pixel Density = 1.00, Graphics Quality = Ultra, Extra Details = 100%.
• World War Z. Resolution 1920 × 1080: DirectX11, Visual Quality Preset = Ultra. Resolution 3840 × 2160: DirectX11, Visual Quality Preset = Ultra.

In all gaming tests, the results are the average number of frames per second, as well as the 0.01-quantile (first percentile) for FPS values. The use of the 0.01-quantile instead of the minimum FPS is due to the desire to clean up the results from random bursts of performance that were provoked by reasons not directly related to the operation of the main components of the platform.

⇡#Performance in complex tests

By tradition, we begin our acquaintance with processors with testing in integrated tests. Benchmark SYSmark 2018 evaluates the weighted average performance of systems not in some selected applications, but in a complex way, simulating scenarios for solving practical problems of one type or another. The assessment uses common office, creative and auxiliary applications: Acrobat Pro DC, Photoshop CC, Lightroom Classic CC, BowPad 2.3, CyberLink PowerDirector 15, FileZilla 3, Chrome 65, Excel 2016, OneNote 2016, Outlook 2016, PowerPoint 2016 and Word 2016 .

It’s worth warning right away that in SYSmark 2018, the performance of AMD processors looks, frankly, so-so. And this is due to the serious dependence of the speed of typical office applications and the responsiveness of the system during multitasking in general on the latencies of the memory subsystem and inter-core communications. And with this, both in the Zen / Zen + microarchitecture and in Zen 2, the situation is not very good. That is why the overall result of the Ryzen 7 3700X in SYSmark 2018 is slightly lower than the Core i9-9900K, and in the Productivity (office work) and Responsiveness (responsiveness) scenarios, it also loses to the Core i7-9700K.

A certain “bright spot” here may be that in a scenario simulating work with creative applications, the Ryzen 7 3700X manages to get ahead of the Core i7-9700K, which shows the good potential of the new CPU in cases where it is faced with multi-threaded and resource-intensive computing workloads.

Complementing the results demonstrated by processors in SYSmark 2018 are performance indicators measured in the 3DMark Time Spy Extreme synthetic gaming test, which is distinguished by high-quality multi-threading optimization and modern instruction sets.

In this test, the Ryzen 7 3700X performs much more confidently. It significantly outperforms its predecessor, with the result that its performance falls in the gap between Intel’s octa-core performance. And for a $329 processor, this is a significant achievement.

⇡#Performance in applications

And now it’s time to look at the most impressive achievements of the Ryzen 7 3700X. Based on the processor’s performance benchmark results in demanding applications, it delivers an average of 25 percent performance gain over the Ryzen 7 2700X, which is made up of both IPC gains, more cache, and higher clock speeds. And as a result, AMD’s junior eight-core processor is able to compete not only with the Core i7-9700K, but also with the faster Intel flagship, the Core i9-9900K.

Most often, the results of the Ryzen 7 3700X fall between the performance of the Core i7-9700K and Core i9-9900K, but you can also see a sufficient number of cases when the lower octa-core Zen 2 microarchitecture outperforms the Core i9-9900K, despite the fact that the Intel processor uses noticeably higher clock speeds. This happens most often in those applications that can clearly benefit from the huge, 32 MB L3 cache of the Ryzen 7 3700X, such as photo processing, archiving, or compiling program code.

At the same time, there are practically no cases when the Ryzen 7 3700X would be slower than the Core i7-9700K. And this means that for workloads, it is now definitely worth choosing mainstream AMD processors, not Intel. In other words, the new Zen 2 microarchitecture is simply ideal for relatively inexpensive computers used in development, design, rendering, video editing and scientific calculations.

Rendering:

Photo processing:

Video work:

Video transcoding:

Compilation:

Archiving:

Encryption:

Chess:

⇡ # Performance in games

⇡ # Tests in FullHD resolution

The biggest question regarding Ryzen 3000 gaming performance is whether those microarchitectural improvements in Zen 2 are enough for the new processors to compete with Intel’s offerings for gaming systems. And the answer to this question is given by the following graph, which we built by averaging the test results in 11 games.

As you can see, the new generation of Ryzen has greatly improved its gaming performance. However, the progress was still not enough for the Ryzen 7 3700X to be able to catch up with the eight-core Intel. In terms of average frame rates, the Core i7-9700K still outperforms the Ryzen 7 3700X by up to 13%.

However, this difference in gaming performance no longer seems dramatic. If the Ryzen 7 2700X was a third worse than Intel’s offerings in games (with a flagship graphics card and in FullHD resolution), now the gap has been reduced by two and a half times. And this at least opens the way for the Ryzen 7 3700X into high-performance gaming systems.

⇡#Tests in 4K resolution

At 4K resolution, where the bulk of the load is shifted to the graphics card, the picture for the Ryzen 7 3700X becomes even more blissful. In fact, this processor ceases to be a significant limiter of gaming performance, and a system based on it and with the flagship GeForce RTX 2080 Ti graphics card is able to show almost the same FPS level as a similar configuration built on eight-core Intel processors. The difference in average speed is no more than 2-3%.

True, it should be borne in mind that, as tests in FullHD showed, the Core i9-9900K and Core i7-9700K processors have the best headroom for gaming performance, which may suddenly be required when more powerful graphics cards appear on the market.

In addition, there are games where the performance gap between the Ryzen 7 3700X and the Core i7-9700K can reach a more significant level. For example, in Assassin’s Creed: Odyssey, GTA V or Total War: Warhammer II, Intel processors are faster than Ryzen 7 3700X, not by 2%, but already by 5%. And if you look at the minimum frame rate, then, as it turns out, the advantage of Coffee Lake Refresh can exceed all 10% even at 4K resolution.

In the end, it remains to be recognized that, although the representative of the Ryzen 3000 family began to look much more attractive in games, it is still impossible to call it an omnivorous and universal gaming solution. Despite noticeable improvements, the new generation of Ryzen still remains a kind of compromise for gaming assemblies, but the scope of applicability of AMD processors in this role will clearly expand.

⇡#Streaming performance

Many gamers choose powerful processors based on the desire to stream. Therefore, we added another game scenario to testing — streaming using the processor. This time, the game Far Cry 5 was used for streaming tests. The popular Open Broadcasting System (OBS) Studio application was responsible for encoding the video stream. It used x264 software encoder. The broadcast was carried out in a resolution of 1920 × 1080 at a frame rate of 60 FPS and a fixed bitrate of 6 Mbps. In the encoding settings, the medium quality settings profile was selected.

There were no complaints about the quality of the outgoing video stream when using only two processors: Core i9-9900K and Ryzen 7 3700X. Both options are completely sufficient for real-time encoding of game content using the medium encoding profile without dropping frames. However, the Intel version is able to provide a noticeably better frame rate on the transmitting side, which can also indirectly affect the quality of the broadcast.

⇡#Power consumption

It would seem that the Ryzen 7 3700X by all accounts should be a very economical processor. First, the underlying semiconductor die is a 7nm CCD chip, and TSMC, when it announced its 7nm process, promised a 40% reduction in power consumption compared to 10nm and a 60% – compared to 14nm chips of similar design. Secondly, AMD has installed a modest 65-watt thermal package for its junior octa-core processor.

However, expectations in terms of energy efficiency turned out to be too high. Due to the fact that in the end the work of technology precision Boost 2 has been reconfigured in favor of maximum frequencies rather than low consumption, no particular economy can be traced in the Ryzen 7 3700X. A system based on this processor in computational algorithms can consume 150 or even 200 watts.

Although in fairness it should be noted that in real tasks the Ryzen 7 3700X naturally consumes less power than AMD’s previous older eight-core, Ryzen 7 2700X, which has a typical heat dissipation of 95 watts. This means that it is also more economical than any eight-core Intel processors with a 95-watt TDP.

At the same time, there is one special case with the consumption of the Ryzen 7 3700X, which should be mentioned separately. This case is a load using AVX2 instructions. As we noted above, their execution in the Zen 2 microarchitecture is twice as fast, so in applications that rely on algorithms rich in vector instructions, the Ryzen 7 3700X catches up with the Ryzen 7 2700X in terms of consumption.

But this is not bad at all, because the difference in performance of the Ryzen 7 2700X and Ryzen 7 3700X in such programs is quite significant in favor of the new processor: to see this, just go back a little and look at the performance of the Ryzen 7 3700X in x264 and x265 encoders.

Moreover, in the process of measuring power consumption in Prime95 29.8 using AVX2 commands, a surprising thing turns out. The Ryzen 7 3700X, compared to the Core i9-9900K, which demonstrates similar performance under AVX2 load, requires one and a half times less electricity. And in general, while the use of AVX2 provokes a tremendous increase in power consumption and heat dissipation in Intel processors, it does not cause anything similar in the new Ryzen 7 3700X. Consumption, of course, is increasing, but on a much more moderate scale. In other words, everything points to the fact that AMD developers managed to come up with some energy-efficient mechanisms for executing vector instructions.

⇡#Conclusions

So far, we have not yet got acquainted with the capabilities of the older representatives of the Ryzen 3000 lineup with 12 and 16 cores, but what we saw today on the example of the younger eight-core Ryzen 7 3700X is really impressive. In just over a year, which has passed since the appearance of the Ryzen processors of the two thousandth series on the market, AMD engineers managed to launch a new project into production, which is radically different for the better from all its predecessors.

In short, in the end, the Ryzen 3000 justified almost all the hopes placed on them, and even more. They seriously increased both single-threaded and multi-threaded performance, became noticeably more economical, and were also able to get rid of a significant part of the «features» of the memory controller.

If you go through the entire list of complaints that we had about the Ryzen processors of the first and second generations, it turns out that AMD was able to delete most of the points from it. In fact, only two particularly annoying moments remain in force: firstly, the connection between the frequency of the Infinity Fabric bus and the frequency of the memory, and, secondly, noticeable delays in the interaction of cores belonging to different CCX complexes. And even then, the Ryzen 3000 implements quite effective measures that allow you to partially mask these problems.

As a result, Zen 2 microarchitecture carriers outperform comparable competitor processors in the vast majority of computational tasks, both multi-threaded and low-threaded, yielding to them only in those cases where data access delays are especially important for performance. More specifically, this means that members of the Ryzen 3000 family will perform exceptionally well when used in creative content creation applications and other resource-intensive professional tasks. The lack of performance can manifest itself in only one case: when it comes to performance in games. But here, the new Ryzen 3000 is no match for its predecessors. Gaming performance has improved significantly, and even in such adverse conditions, the new AMD processors look far from hopeless against the background of Intel Core.

In addition, according to the established tradition, AMD generously compensates for all shortcomings in consumer qualities with prices, which can be traced this time as well. The eight-core Ryzen 7 3700X tested in this review is just a vivid illustration of this thesis. The $329 listing price makes it a cheaper alternative to the Core i7-9700K. And this means that if we are talking about finding a platform for work or for versatile applications that include resource-intensive loads, then there are no objective reasons for choosing the Intel option. In other words, the Ryzen 7 3700X pushes the Core i7-9700K, and even the more expensive Core i9-9900K, into the pure gaming segment.

Of course, it cannot be said that the only thing that caused us to feel slight dissatisfaction with the Ryzen 7 3700X when we met was its gaming performance. In fact, you can get upset for many other reasons, for example, because of the disappeared overclocking potential or lost energy efficiency somewhere. In addition, it is impossible not to mention that AMD once again dooms the first Ryzen 3000 buyers to the role of beta testers, since the Socket AM4 platform in its current state is teeming with various flaws and is not yet ready to accept new CPUs. However, I really don’t want to talk about it. In the end, AMD’s product this time really turned out to be very worthy, and it’s simply inappropriate to grumble today.