As you already understood, this was an April Fool's prank, but we really hope that AMD will keep its word and the results of the final sample will not differ much from those indicated in the review, because all the slides are authentic, that is, AMD really promised 40% IPS for AMD Zen in comparison with the previous generation.
Surely many people know that as part of large exhibitions, closed presentations of certain products are held, where not all guests are allowed and only by invitation. One of them at CeBIT 2016 was organized by AMD, demonstrating its new products to key partners and investors. As we were told, one of the highlights of this closed presentation was an engineering sample of a new desktop processor with a 14 nm microarchitecture. We hope that within the framework of the upcoming Computex 2016, AMD will have the opportunity to demonstrate a full-fledged final sample, exactly as planned.
Therefore, when we were asked to postpone all our current tests and for a couple of hours have an engineering sample of the AMD Zen processor at our disposal for testing (albeit with a number of restrictions), we did not hesitate for a minute to answer - after all, the case is truly unique. And the restrictions turned out to be quite mild: do not show the back side of the processor itself and the motherboard used, and also do not try to overclock. Otherwise, there were no restrictions on the benchmarks used.
Traditionally, we begin our review of a processor with its specification and a brief analysis of innovations, if we are talking about a new generation. IN in this case the specification table will consist only of the information provided to us, and the microarchitecture review will consist of the crumbs of information that we found on the Internet, because a colorful and informative presentation on AMD Zen is not yet ready from AMD itself. So let's begin.
Specification:
|
AMD Zen engineering sample |
|
|
Market segment |
Desktop systems |
|
CPU socket |
|
|
Manufacturing process, nm |
|
|
Microarchitecture |
|
|
Number of physical cores/threads |
|
|
Nominal clock frequency, MHz |
|
|
L1 cache |
Unknown |
|
L2 cache, KB |
|
|
L3 cache, MB |
|
|
Supported RAM |
DDR4-2400 MHz |
|
TDP indicator, W |
SMT vs SMT: return to the classics
If you follow the development of the situation in the traditional processor market over the past 12 years, you can see that the turning point came in the second quarter of 2006. According to the results of the first, AMD's market share rose to 48.4%, and Intel dropped to 51.6%. But then Intel introduced its successful and famous Intel Core microarchitecture, the successors of which to this day allow it to dominate the traditional market computer systems. AMD at that time had a pretty good, but still not competitive enough AMD K8 microarchitecture. In September 2007, the AMD K10 microarchitecture was released, but it did not help AMD regain its previously lost positions. Nevertheless, work was already in full swing on an update - AMD Bulldozer, which was supposed to mark a transition to a qualitatively new level and become a worthy answer to Intel Westmere and the future Intel Sandy Bridge. The presentation of the AMD Scorpius platform and the first processors in the line took place in October 2011. But already the first tests were a real disappointment for the public - they not only did not bring a significant increase in performance, but in some benchmarks they were even slightly inferior to the previous generation of AMD CPUs. What can we say about the new Intel processors?
The transition to CMT (Clustered Multi-Thread) technology played a key role in this fiasco. Without going into a deep analysis, we will only briefly recall that along with the AMD Bulldozer microarchitecture, the concept of a processor module was introduced, which combines two integer calculation blocks and one real calculation block using SMT (Simultaneous Multithreading) technology to simultaneously process two threads. That is, from the point of view of integer calculations, there are two physical processor cores in one module, and from the point of view of real ones, there is one physical core and two virtual ones. In turn, Intel uses exclusively the SMT approach: there is a full-fledged physical core with the required number of integer and real calculation blocks, and SMT technology is applied to it for parallel processing of two threads.
AMD's idea was not bad, but the company lost sight of a very important point - the need to optimize the program code of specific applications for a multi-threaded modular system. Indeed, in 2011, most programs worked in single-threaded mode, so it was more important for them to have one full-fledged physical core in the processor than four modules. Subsequently, AMD worked closely with Microsoft to optimize the program code of the Windows OS family and with other developers to actively integrate the idea of parallel computing, but optimizing the program code required time and money, and AMD was losing customers and financial resources.
Realizing the scale of the situation, the company's management decided to create a completely new microarchitecture. This process takes several years, during which AMD could only slightly improve the AMD Bulldozer concept. Jim Keller, a very authoritative and respected specialist in the industry, was invited to the post of lead architect. It was he who was involved in the creation of the AMD K7 microarchitecture and worked as the lead architect in the creation of AMD K8, which was able to bring AMD as close as possible to Intel in the first quarter of 2006. After completing work on AMD K8, Jim Keller joined Apple, and under his leadership the legendary Apple A4 and Apple A5 chips were released.
From 2012 to 2015, Jim Keller and a team of engineers worked on creating the AMD Zen microarchitecture, which was announced to the general public only in the second half of 2015. The first thing that was emphasized during the announcement was the abandonment of CMT and the transition to full-fledged SMT. This means that AMD Zen will use separate physical cores with the necessary set of all building blocks: 4 ALUs for integer calculations, 4 FPUs with a 128-bit bus (combined into two 256-bit FMAC modules) for real-valued calculations, and 4 decoders. And thanks to the SMT approach, each core will be able to process two data streams in parallel (similar to Intel Hyper-Threading technology). The maximum number of physical cores for desktop processors will reach 8, and for server processors - 32.
It is also known from unofficial sources that each core uses 512 KB of L2 cache, and every 4 cores share a common 8 MB of L3 cache. The optimization of the AMD Zen microarchitecture for popular modern compilers was also discussed, that is, the new processors will no longer require any optimization of the program code on the part of developers, but can immediately offer an optimal level of performance. As a result, such an important indicator as IPS (Instructions per Clock) should increase by 40%. I wonder if we can get a similar increase?
From theory to practice
Now let's move on to a test sample of a 14nm processor with AMD Zen microarchitecture. At the time of review, the CPU-Z utility did not officially support these solutions, so to analyze the data we used AIDA64, to which support for AMD Zen was added from version .
The nominal frequency of the engineering sample turned out to be 3.3 GHz. It is quite possible that in the final version the frequency will increase slightly (within 100 MHz), but you should not expect a more significant increase - after all, 8 cores and 16 threads cannot operate at higher speeds while maintaining a 95-watt thermal package. By the way, it was the use of the energy-efficient 14-nm FinFET LPP process technology that made it possible to achieve such indicators. For contrast, remember that the 22nm 8-core processor has a base frequency of 3.0 GHz and a TDP of 140 W.
To cool the AMD Zen engineering sample, we used a cooler. Which is capable of handling 125-watt processors. As you can see, the temperature remained at 57°C. We do not know the critical value of this parameter for AMD Zen, but the processor itself worked stably, without any errors.
The exact structure of the cache memory could not be determined, since CPU-Z does not yet know about the existence of AMD Zen. Therefore, we repeat that according to preliminary data, we have 512 KB of L2 cache per core and 8 MB of L3 for every four processor cores. That is, the total L3 cache size reaches 16 MB. If we continue the comparison with the same Intel Core i7-5960X Extreme Edition, we see a double increase in L2 cache memory (512 KB versus 256 KB), but a lag in L3 volume (16 MB versus 20 MB).
Built-in controller random access memory supports work with DDR4-2400 MHz modules. There was information that when overclocked, the memory frequency could reach DDR4-2933 MHz, but we were forbidden to test such a theory.
The AMD Zen engineering sample does not have integrated graphics. It won't be in the final version either. However, next year they promise to transfer the new generation of APUs to the 14-nm AMD Zen microarchitecture, adding a 14-nm iGPU AMD series Polaris.
Testing
During testing we used Processor Test Stand No. 2
| Motherboards (AMD) | ASUS F1A75-V PRO (AMD A75, Socket FM1, DDR3, ATX), GIGABYTE GA-F2A75-D3H (AMD A75, Socket FM2, DDR3, ATX), ASUS SABERTOOTH 990FX (AMD 990FX, Socket AM3+, DDR3, ATX) |
| Motherboards (AMD) | ASUS SABERTOOTH 990FX R2.0 (AMD 990FX, Socket AM3+, DDR3, ATX), ASRock Fatal1ty FM2A88X+ Killer (AMD A88X, Socket FM2+, DDR3, ATX) |
| Motherboards (Intel) | ASUS P8Z77-V PRO/THUNDERBOLT (Intel Z77, Socket LGA1155, DDR3, ATX), ASUS P9X79 PRO (Intel X79, Socket LGA2011, DDR3, ATX), ASRock Z87M OC Formula (Intel Z87, Socket LGA1150, DDR3, mATX) |
| Motherboards (Intel) | ASUS MAXIMUS VIII RANGER (Intel Z170, Socket LGA1151, DDR4, ATX) / ASRock Fatal1ty Z97X Killer (Intel Z97, Socket LGA1150, DDR3, mATX), ASUS RAMPAGE V EXTREME (Intel X99, Socket LGA2011-v3, DDR4, E-ATX ) |
| Coolers | Scythe Mugen 3 (Socket LGA1150/1155/1366, AMD Socket AM3+/FM1/ FM2/FM2+), ZALMAN CNPS12X (Socket LGA2011), Noctua NH-U14S (LGA2011-3) |
| RAM | 2 x 4 GB DDR3-2400 TwinMOS TwiSTER 9DHCGN4B-HAWP, 4 x 4 GB DDR4-3000 Kingston HyperX Predator HX430C15PBK4/16 (Socket LGA2011-v3) |
| Video card | AMD Radeon HD 7970 3 GB GDDR5, ASUS GeForce GTX 980 STRIX OC 4 GB GDDR5 (GPU-1178 MHz / RAM-1279 MHz) |
| HDD | Western Digital Caviar Blue WD10EALX (1 TB, SATA 6 Gb/s, NCQ), Seagate Enterprise Capacity 3.5 HDD v4 (ST6000NM0024, 6 TB, SATA 6 Gb/s) |
| power unit | Seasonic X-660, 660 W, Active PFC, 80 PLUS Gold, 120 mm fan |
| operating system | Microsoft Windows 8.1 64-bit |
Select what you want to compare AMD Zen Eng Sample with
What can we expect from the company in 2017?
Some time ago, AMD shared with the general public another piece of data about the new Zen microarchitecture, as well as the AM4 platform, which (together with new processors and APUs) should become the company’s main product for the desktop market starting next year. It is clear that the preliminary information is not exhaustive, but it is quite interesting, since it allows you to roughly understand what you should expect from new products (and what you should not). This was the reason for writing this material, which is dedicated not to microarchitectural subtleties (certainly important, but not all), but, let’s say, to the consumer characteristics of the new platform.
Current issues
As we wrote almost two years ago, the situation with AMD desktop platforms has looked a little strange over the past few years. In fact, the main events took place in the field of APUs (as the company calls processors with integrated graphics), where two and a half platforms have changed since 2011: FM1, FM2 and FM2+, compatible with the latter from top to bottom. However, all of the listed solutions (even the FM1 platform, which did not stay too long on the market) can be considered modern: high degree integration allows you to create complete systems using literally a couple of chips - the processor itself (most of which are equipped with GPUs that are excellent by the standards of integrated solutions) and a chipset. The line of chipsets also meets modern requirements - in terms of integration functionality AMD was often ahead of Intel, being the first to provide its chips with built-in support for USB 3.0 and speeds of 6 Gbps for all SATA ports, for example. The only thing that prevented the wide expansion of solutions for this platform was the relatively low performance and high power consumption of the APU processor part in comparison with competing solutions. Higher performance could be obtained by choosing solutions for the AM3+ platform, which essentially dates back to the platforms of the beginning of the century. And the multi-module processors themselves have not been significantly updated since 2012, so they could only be sold due to low prices at a relatively high cost due to the use of the already outdated 32 nm process technology. The latter, to some extent, also applied to APUs, which during their existence “moved” from the mentioned standards to only 28 nm, which is also not the peak of technology for a long time - in many ways this is what caused the mentioned problems with energy consumption.
It is worth noting that the company never considered this state of affairs to be “normal”: the unification of platforms was initially planned for 2012. However, in practice this did not happen, so a kind of “sitting on two chairs” continues to this day. Thus, in essence, both processors and AMD platforms are now outdated, so the situation needs to be changed radically. This is what the company plans to do.
AM4: finally a single platform
AMD has fully confirmed the existing assumptions about the characteristics of the new platform, even “with a slide”. In particular, to key features AM4 company attributes the following:
- DDR4 memory
- Full PCIe 3.0 support
- USB 3.1 (“full”, i.e. Gen2 with speeds up to 10 Gbps)
- NVMe and SATA Express
Concerning last point, then, in principle, serious hardware modifications were not required for its implementation: it is possible within the framework of existing platforms. In particular, many motherboard manufacturers have even updated their range of models with AM3+, allowing them to boot from NVMe drives. More important for the full functioning of NVMe drives at maximum speed is support for PCIe 3.0, which was not available within AM3+ at all, and APUs for FM2+ supported only 24 lanes of this interface, some of which “went” to communicate with the chipset, and 16 could be required by the video card. In addition, as mentioned above, there were no high-performance processors for FM2+, so the platform has long been firmly established in the budget sector, where the NVMe protocol is not very relevant (simply because so far all drives that support it are exclusively “non-budget”). According to plans, AM4 should become a solution for all market segments, so this may become necessary for it - especially considering AMD’s desire to create “long-lived” platforms, which is highly valued by many users. Exactly the same applies to USB 3.1 support: for now it is not necessary, but it may come in handy in the future. Again, as mentioned above, AMD implemented the previous version of the standard in chipsets a year earlier than Intel, so it’s logical to expect the same for new version USB.
Mastering DDR4 is a long-awaited step, since the performance of integrated GPUs is highly dependent on memory bandwidth. Previously, this problem had to be solved by increasing DDR3 frequencies, but this approach, to put it mildly, is not ideal in terms of price and power consumption of modules. Actually, this is why there have been talks about introducing DDR4 support in AMD APUs since 2013 (at that time there were a lot of assumptions about two options in the upcoming Kaveri), but for a long time new memory modules were too expensive for use in mass systems. On this moment DDR4 shipments are already outpacing DDR3, so prices have leveled off - with a trend in favor of DDR4. In general, the time has come to say goodbye to the old standards, and, apparently, AMD plans to do this more abruptly than Intel - which, we recall, has not yet completely abandoned DDR3. On the other hand, the last major update to LGA115x was last year, and the most interesting products for AM4 will appear next year, so this difference in approaches is quite understandable.
Bristol Ridge: an interim solution
However, the “running-in” of the platform has almost begun: as expected, a number of processors for it have been released right now and are already being shipped to large manufacturers. All of them still belong to the budget segment, so the company has so far squeezed the most functional of the chipsets (X380), supplying only a couple of inexpensive modifications - A320 and B350. However, in practice, these will be enough for many. What they don't have is PCIe 3.0 support - only 4 or 6 PCIe 2.0 lanes, respectively. On the other hand, 10 PCIe 3.0 lanes (not counting those needed for communication with the chipset) are supported by the current processors/APUs themselves, and the presence of powerful (for solutions of this class) graphics in these APUs in an inexpensive computer will definitely leave PCIe processor lines free for peripherals.
In general, in fact, one can observe the unification of mobile and desktop solutions: APUs of the Bristol Ridge family are the heirs of the Carrizo we are already familiar with. In addition to the mentioned 10 PCIe 3.0 lanes (x8+x1+x1, the last two can be simultaneously “given” to an NVMe drive), they themselves support 4 USB 3.0 ports (aka USB 3.1 Gen1) and 2 SATA600 ports. The use of the lower-end A320 chipset adds to the above a USB 3.1 connector (full-speed, as noted above), 2 USB 3.0 ports, 6 USB 2.0 ports, 4 PCIe 2.0 lines, 2 SATA600 ports and 1 SATA Express connector (which can be used as a SATA pair ). The B350 has similar functionality, but adds 1 more USB 3.1 port and 2 PCIe 2.0 lines. In addition, according to good tradition, everything AMD solutions support the creation of RAID arrays of levels 0, 1 and 10.
How does this compare to Intel's budget offerings like the H110 and B150? To simplify understanding, we will collect the characteristics of the platforms in a table, adding to it the mass-produced A78 for FM2+, which is leaving the market.
| Chipset | AMD A78 | AMD A320 | AMD B350 | Intel H110 | Intel B150 |
| PCIe 3.0 lanes (total) | 8/16 | 10 | 10 | 16 | 24 |
| PCIe 2.0 lanes | 4 | 4 | 6 | 6 | 0 |
| SATA600 ports | 6 | until 6 | until 6 | 4 | until 6 |
| RAID 0/1/10 | Yes | Yes | Yes | No | No |
| SATA Express ports | 0 | 1 | 1 | 0 | 0 |
| USB 3.1 ports | 0 | 1 | 2 | 0 | 0 |
| USB 3.0 ports | 4 | 6 | 6 | 4 | 6 |
| USB 2.0 ports | 14 | 6 | 6 | 6 | 6 |
So, the only formal weak point of the new platform is the number of PCIe 3.0 lanes provided by the processor: only 10 versus the usual 16 in the mass segment. But this weak point is only Bye- it’s just that at the moment there are no other APU models, but they will appear in the future. In the end, solutions based on FM2+ (A78) may not have PCIe 3.0 lanes at all - if you install an FM2 processor in the board, which only supported PCIe 2.0. But Intel platforms have a different problem: all processors for LGA1151 support PCIe 3.0 x16, but on boards with budget chipsets this line configuration will be the only one - these lines are not allowed to be “split” into slots/devices. AMD adheres to a different practice, so that in a system with an A320 you can, for example, “drive” two NVMe drives on PCIe 3.0 - but in a system with an H110 you cannot (however, PCIe 3.0 x2 is equal in bandwidth to PCIe 2.0 x4, but in many ways Is it possible for inexpensive H110 boards to implement at least such a slot?). To what extent this (as well as support for SATA Express or RAID arrays) is in demand in inexpensive systems is a separate question. But the fact remains: in fact, even the lowest versions of the new platform are comparable in functionality to older Intel solutions.
As for the capabilities of connecting external peripherals, chipsets for FM2+ continue to hold the record in terms of the total number of USB ports. But this record is purely theoretical - in fact, so much USB 2.0 in final solutions is simply not in demand. But four high-speed USB ports are sometimes not enough, which also affects the Intel H110. At the same time, the youngest chipset for AM4 supports seven USB 3.0 ports (one of which is actually USB 3.1, which for now, as mentioned above, is mainly a foundation for the future, but at USB 3.0 speed this port can be used now) - even more than B150. Perhaps, in the “200th” series of chipsets, Intel will “tweak” the younger modifications, but it is not there yet, and the A320 and B350 are already being shipped to manufacturers.
The development of compact computers based on AMD processors should sparkle with new colors, since some of the functionality of traditional chipsets has already been transferred to the processors themselves, which to some extent makes AM4 similar not only to FM2+ or AM3+, but also to AM1. In AM1, however, the functionality of the SoC was very limited, and there were no possibilities for its expansion, but now this problem has been removed. More precisely, it was filmed in laptop Carrizo a year ago, and it is not surprising that when developing a new desktop platform, these achievements were taken into account and inherited. What does this mean in practice? For example, without any special difficulties, you can produce Mini-STX boards with a replaceable processor, but by “saving” on the chipset chip - four USB 3.0 ports and a pair of SATA600 (one of which, in combination with PCIe 3.0 x4, can be reasonably allocated to the M. 2) there's enough there. There were difficulties with this before, but not anymore.
| CPU | AMD A12-9800 | AMD A12-9800E | AMD A10-9700 | AMD A10-9700E | AMD A8-9600 | AMD A6-9500 | AMD A6-9500E | AMD Athlon X4 950 |
| Production technology | 28 nm | |||||||
| Core frequency std/max, GHz | 3,8/4,2 | 3,1/3,8 | 3,5/3,8 | 3,1/3,5 | 3,1/3,4 | 3,5/3,8 | 3,0/3,4 | 3,5/3,8 |
| Number of modules/calculation threads | 2/4 | 2/4 | 2/4 | 2/4 | 2/4 | 1/2 | 1/2 | 2/4 |
| L1 cache (total), I/D, KB | 192/128 | 192/128 | 192/128 | 192/128 | 192/128 | 96/64 | 96/64 | 192/128 |
| L2 cache, KB | 2×1024 | 2×1024 | 2×1024 | 2×1024 | 2×1024 | 1×1024 | 1×1024 | 2×1024 |
| RAM | 2×DDR4-2400 | |||||||
| TDP, W | 65 | 35 | 65 | 35 | 65 | 65 | 35 | 65 |
| Graphic arts | Radeon R7 | Radeon R7 | Radeon R7 | Radeon R7 | Radeon R7 | Radeon R5 | Radeon R5 | - |
| Number of GPs | 512 | 512 | 384 | 384 | 384 | 384 | 384 | - |
| Frequency std/max, MHz | 1108 | 900 | 1029 | 847 | 900 | 1029 | 800 | - |
But why, with all these interesting features, are we inclined to consider the current implementation of the platform an intermediate solution? The fact is that the processors currently available for it are very limited. AMD, of course, praises the “seventh generation” APU, but the same was said about previous models. But in practice, this is only a further development of the same modular architecture that debuted back in 2011, and the same 28 nm process technology, used since 2014. Yes, as our tests have shown, Carrizo processors are often (thanks to optimizations) faster than Kaveri processors running at a higher clock frequency, and support for DDR4 memory should give them a little boost. The integrated GPU was previously one of the best in its class, and since 2015 it has received an updated video processing unit with hardware support for VP9 and H.265/HEVC with resolutions up to 4K. All this is true - but it only leads to evolutionary changes that do not fundamentally change the class of the solution. Thus, the only Athlon X4 for the new platform at the moment, the model with index 950, is identical in everything except the type of RAM to the Athlon X4 845 for FM2+, and more or less close analogues can be found for other new processors. Therefore, the real launch of the AM4 platform is expected only next year - at least if AMD’s plans are fulfilled.
Zen: what's new?
So, what problems did the company face? The primary controversial point of the developed modular architecture was the modules themselves: to save the transistor budget, the pair of “x86 cores” included in them depend on each other, since they separate some blocks. In particular, in the first implementations even the instruction decoder and instruction cache were unified. The second weak point is the memory system. At the time of the development of the first processors, it was possible to make a fast second-level cache, but L3 remained external to the main part of the processor, so it worked asynchronously with it and at lower clock frequencies. As a result, in higher configurations of the FX family of processors, the total L2 capacity turned out to be equal to L3, which forced AMD to continue using an exclusive cache memory architecture. It worked great in the days of single-core processors, but made it difficult to exchange data between computational threads in multi-core ones, complicating the algorithms: if something is not in L3, it may be in L2 of one of the modules, or maybe only in memory. And even a single L2 for a pair of cores, so convenient for the Core 2 Duo, could not be used for synchronization: the greatest efficiency was demonstrated by the module executing only one stream of commands, i.e., loading the “second halves” (in fact, a smaller part of them) work only made sense if there was too much of it, but not on the usual two to four threads for massive loads.
And in APUs, most of the chip was occupied by the graphics core, so these models were left without a single cache memory, even a slow one, since otherwise the processor would have been too large. In fact, when using the same production standards, APUs competed in cost with older quad-core models of the mainstream Intel processor line, and older processors with four modules turned out to be even more expensive. But at the same time, one could talk about competition in terms of performance only by comparing four AMD modules with four Intel cores- only one SIMD block per module added fuel to the fire. Wherein Intel processors and in themselves were cheaper to produce, but due to the features of the platforms they cost significantly less. APUs “fought” only with very cheap dual-core Intel processors, and even this was done with varying degrees of success. Of course, they had an advantage in graphics performance, but it was not always in demand.
What is changing in the new generation (as we promised - in simple language, without going into technical jungle)? The “base element” of Zen is somewhat reminiscent of the dual-module processor of the previous architecture, but with significant modifications. Firstly, it includes not four pairwise “x86 cores”, but four full-fledged and independent cores - independent even in terms of second-level cache memory, the total capacity of which has been halved, but now each core has its own L2 (and, of course, its own instruction decoder along with instruction cache memory). Secondly, third-level cache memory has become an integral component of such a “building block”. Apparently, it will work significantly faster than its predecessors, and its capacity is 8 MB. Thirdly, and importantly, AMD also managed to implement symmetric multithreading technology, so that each core can execute commands from not one, but two threads.
In fact, as you can see, in the “basic” version Zen strongly resembles the top mainstream Intel processors, i.e. quad-core Core i7. Moreover, such a “module” in the second half of next year will also be used in APUs, which currently only have, let us remind you, two “old-style” modules, and without any third-level cache at all. The graphics core may not be on par with Intel's top solutions (especially those equipped with L4 cache - AMD has not yet promised anything like that), but it will be more productive than mass-produced Intel integrated graphics. Moreover, judging by the available data on the internal organization of processors, the company will be able to develop a budget modification with a pair of cores and L3 reduced to 4 MB, i.e., release direct competitors for a variety of Core i3 and other dual-core processors (especially mobile ones). Now only dual-module (in AMD terminology - “quad-core”) processors can compete with them, and in the future “regular” dual-core processors will do the same.
However, it cannot be said that the company has completely achieved “core parity”. In particular, blocks for working with floating point numbers and other SIMD instructions have changed less than we would like. They do not have normal support for working with 256-bit vectors, i.e. you cannot expect good results with AVX2 code. On the other hand, at the moment it is premature to say anything about performance - the new microarchitecture will debut in finished products only next year. Then there will be complete clarity with their clock frequencies, prices, and performance in real tasks. For now, we can only evaluate AMD's plans.
And in them there was also a place for lovers of high processor performance, since there will be at least two options for the layout of finished products (and if we take into account the possibility of releasing dual-core models that can easily find their place in the budget segment, then three): in addition to the APU, where, as already As mentioned above, one quad-core Zen “module” will be adjacent to the GPU; it is also planned to release “pure” CPUs - with two modules. That is, such solutions will have 8 cores capable of simultaneously performing 16 computation threads and equipped with a third-level cache with a capacity of 16 MB. With L3 it is not completely clear whether it will be a single volume accessible to all cores of a “composite” processor, or two separate blocks (which is inherent in “gluing”), but the capacity will be exactly the same. At the same time, top processors will remain compatible with the same AM4 platform, which is an important competitive advantage over Intel processors for LGA2011-3 and their successors, which are mechanically incompatible with the mass line. Yes, of course, what was said above about the performance of vector instructions will be true, and the memory controller of these new models will remain two-channel rather than four-channel, but the latter also has its advantages: the boards will be cheaper. Moreover, these will be the same boards as for low-cost APUs, i.e. the long-awaited single AMD platform will probably be able to be used even more widely than Intel LGA115x. And if the company also manages to “fix” it for five years (implementing at least top-down compatibility), turning it into a “long-liver” of the AM3 class, so much the better for many consumers.
A logical question arises, of course: if all the changes are so logical and expected, then why did the “waiting” last so long? After all, in a good way, such devices are needed “yesterday”, and the company plans to deliver them only “tomorrow”. There is a problem, but it does not affect development itself - only production. In fact, all that was available to AMD until recently was a 32 nm process technology, which is only sufficient for FX. At best, it will reach the level of Intel Sandy Bridge, which is also more than five years old. The latest APU models, however, use 28 nm standards, but this is not much better than 32 nm. Therefore, a “big leap” is planned in production - a transition to the 14 nm process technology. The transition will take place with some lag behind Intel (which has been using this technical process for two years), but understandable and explainable. In general, it was impossible to make such processors without mastering new production standards - and mastering them takes time. We want to believe that AMD will succeed.
Total
So what do we get? First of all - finally! - transition to a single platform, which has not happened for five years. Moreover, in this case, we can talk about a “big leap”: according to plans, AM4 should be more universal than Intel LGA115x. Secondly, a significant change in microarchitecture - with an increase in the performance and overall efficiency of processors based on it. Thirdly, a sharp improvement in production standards, which is good in itself, and without which such changes would be impossible. That is, as you can see, AMD plans to eliminate all the shortcomings of today's mass-produced systems in one fell swoop. Will it work? Only practice will show this - for now we can only evaluate plans and preliminary information. However, in some way The AM4 platform already exists, and in its price segment it has a number of advantages over competing developments. Basically, they are inherited from their predecessors (this is not surprising - it is difficult to call the APUs currently being released “new”), but with the addition (at least potentially) of upgradability and a longer life cycle. And we will get the final answer to the question of how successful the transition will be next year. I would like to believe that the answer will be positive - at least it’s more interesting :)
AMD has long promised that after the launch of a complete model range Ryzen for desktop systems will be followed by Ryzen Mobile - mobile versions of the new processors. Such processors have long been discussed under the code name Raven Ridge, and according to preliminary data, they will receive four Zen computing cores and a graphics core with Vega architecture. Now, thanks to the Ashes of Singularity benchmark database, real confirmation has been received not only of the existence of such a development, but also that it is in a stage close to the final one. Moreover, the recorded results reveal the name of the future APU: it is called Ryzen 5 2500U.
The AMD Ryzen 5 2500U is likely to be one of the older models in the promising Raven Ridge APU family, aimed at light and thin laptops. If the benchmark data is correct, then such a processor will have four cores with support for SMT (multi-threading) technology and an AMD 1500 Graphics core belonging to the Vega generation. In other words, the two thousandth series in the Ryzen nomenclature will be allocated to the company’s APUs based on the Zen microarchitecture.
Previously, AMD had already confirmed that Raven Ridge processors will indeed receive a graphics core based on the Vega architecture. Also, according to official information, the first laptops with processors of the Ryzen Mobile family will appear on store shelves before the end of the year. Pilot models mobile computers based on new AMD processors, aimed at the mass segment, will likely be announced at the end of the third quarter. Similar products for the corporate market are planned for the first half of 2018.
Compared to the previous generation of APUs, Bristol Ridge, AMD's future hybrid processors with the new architecture will have to offer at least a 50% increase in computing performance and at least a 40% increase in graphics performance while reducing power consumption by more than half.
Unfortunately, the speed indicators of the future APU that appeared in the Ashes of Singularity benchmark database are not yet relevant and do not reflect the real performance of Raven Ridge.
IntroductionOver the past few years, AMD has lost almost all of its previously won positions in the processor market for desktop computers. With the Bulldozer family of cores, the company was stuck in a world of 32- and 28-nm planar transistor chips, while Intel repeatedly made architectural improvements, moved to 3D transistors, and introduced 22- and 14-nm manufacturing processes. As a result, AMD simply had no offers for high-performance computers left in its assortment, and Intel was, in fact, able to occupy a monopoly position. But fortunately, AMD decided not to put up with the current situation and has devoted the last few years to working on a new processor design - the Zen microarchitecture. It promises everything that enthusiasts would like to see in a modern processor: high specificity, good energy efficiency, modern technology production and attractive price. AMD Ryzen is the first processor on the new microarchitecture, and if the developers really delivered on all their promises, then today we will see AMD’s triumphant return to the market.
Zen is a huge step forward compared to past AMD microarchitectures. This is not a further development of Bulldozer, but a completely new and independent project in which it was possible to achieve an unprecedented increase in efficiency. Based on the results of the work, AMD speaks of a 52 percent increase in IPC (the number of instructions executed per clock cycle) compared to the Excavator microarchitecture. In addition, Ryzen, for the first time for AMD, introduces support for SMT (Simultaneous Multi Threading) technology, which allows the execution of two computing threads on one core. At the same time, Ryzen is also the first AMD processor released using the modern 14-nm process technology using FinFET transistors, which helps to conquer high frequencies with good energy efficiency. Another important change is the move to a more modern platform, which is designed to work with dual-channel DDR4 SDRAM.
The Ryzen 7 processor line that AMD is launching today includes three eight-core processors with prices ranging from $330 to $500. They are all similar in basic characteristics, but differ in frequencies. We managed to get the average model in the family for testing, the four hundred dollar Ryzen 7 1700X, which is going to compete with the Core i7-6800K or Core i7-7700K. Assemblies based on new AMD processors are good because motherboards with the required Socket AM4 socket are noticeably cheaper than motherboards for flagship Intel processors, and therefore a configuration based on the Ryzen 7 1700X can indeed become a very attractive option for a desktop personal computer. The main thing is that everything that AMD managed to promise regarding performance and other consumer qualities really comes true.
In other words, today we can witness the most ambitious event in the processor market in the last five years. Real competition may indeed return to the field of desktop processors, and this will be quite capable of pushing forward the noticeably stalled progress. Therefore, we will not postpone the most interesting things until later, but will immediately move on to the technical details, and then to the tests.
Zen microarchitecture: in brief
To understand the ideas behind the new processor design, you need to know that AMD engineers focused on four main aspects when developing the Zen microarchitecture. Firstly, performance. Engineers tried not only to achieve significant improvements in the execution speed of single-threaded workloads, but also sought to increase the parallelism of the architecture whenever possible. Secondly, throughput. The new processors have significantly improved cache memory and prefetch algorithms, and the execution pipeline has been redesigned to avoid bottlenecks and downtime. Thirdly, efficiency. Optimizing the specific output per watt expended was another important priority. Zen uses all AMD's existing developments aimed at managing power in the active state and in idle state, and also takes advantage of all the advantages that the 14-nm process technology with FinFET transistors provides. And fourthly, scalability. New Ryzen processors have a modular design, the main building block of which is the quad-core CCX (Core Complex) unit. These blocks are connected together by the new high-speed Infinity Fabric bus, which makes Zen a design that can be implemented in processors of varying complexity and different purposes.Let's look at all the listed features in a little more detail.
From a performance standpoint, the Zen microarchitecture makes what the company calls a “quantum leap” in instruction execution speed over previous designs. This is primarily due to the fact that the Zen cores no longer share any resources with each other, as was the case in Bulldozer, they are completely independent and also support SMT technology, which allows two threads to be executed on one core simultaneously (analogous to Hyper-Threading) . In addition, each core received its own micro-operation cache, which significantly reduces the overhead of decoding instructions, and a completely redesigned fast first-level cache with write back and low power consumption, each core has its own FPU unit and dedicated L2 cache, as well as a lot of other optimizations.
With a 75 percent increase in the size of the scheduler window, overall schedulers can send 1.5 times as many instructions for execution as they could in Excavator cores. At the same time, the decoder is expanded by at least one and a half times, thanks to which Zen can send significantly more work to their actuators. In addition, Zen introduced a micro-op cache, which allows the processor to avoid repeated calls to the L2 and L3 cache and repeated decoding of instructions when working with repeated sections of code. The branch prediction scheme has changed significantly; it now uses hardware neural network, which significantly increases the percentage of correctly taken branches. Plus, full utilization of all available resources is facilitated by SMT support, which allows applications that support parallel computing to create twice as many threads.
A productive engine always needs adequate fuel supply, and the Zen microarchitecture pays a lot of attention to this aspect. Therefore, you should not be surprised that the cache memory hierarchy has changed somewhat. The L1 instruction cache has increased to 64 KB, and the L1 data cache has become a write-back algorithm. The L2 cache became individual for each core with a volume of 512 KB. And the L3 cache received a volume of 8 MB for every four cores, for which it is shared within the Core Complex. Featuring intelligent prefetching algorithms, the new caching system can feed compute cores up to five times more data than Excavator.
The 14nm process technology also plays an important role in the implementation of the Zen architecture. For the physical implementation of Ryzen processors, AMD chose the GlobalFoundries process technology, which is focused on high-density designs. This made it possible to ensure that the Ryzen core has a relatively small area, operates at fairly low supply voltages, and ultimately provides a favorable relationship between power consumption and performance. In addition, all the company’s past developments aimed at increasing the energy efficiency of the CPU have been used in Zen: dynamic power supply and shutdown of various processor nodes, dynamic frequency changes. Solutions aimed at saving energy can also be found directly in the microarchitecture. This is partly helped by the micro-op cache, and in addition, the CPU manager uses a special stack mechanism to generate reusable addresses.
Thanks to optimizations of this kind, the Zen microarchitecture has a very wide range of applicability; in the future, it should become the basis for the entire family of AMD processor products: for laptops, desktops and servers.
Zen's scalability is partly based on the fact that the processors are assembled from CCX building blocks that combine 4 cores and are capable of executing 8 threads. Each CCX has 512 KB of L2 cache per core and a shared 8 MB of L3 cache. The current Ryzen 7 processors that AMD is introducing today are assembled from two CCXs, and respectively receive 8 cores and 16 threads. The CCXs are connected to each other by a special Infinity Fabric bus.
This stacked Zen design will allow AMD in the future to produce processors with different numbers of cores and threads, different amounts of cache memory, aimed at different applications and market segments.
A significant role in this is played by the Infinity Fabric bus, which is based on HyperTransport and allows you to quickly and with minimal effort assemble processor chips of various configurations. High throughput and traffic prioritization make Infinity Fabric well suited for this role. The bus copes without problems with data transfer between CCX, system memory and other controllers that are presented in the Ryzen processor core. In addition, the parameters of individual CCXs are also managed through Infinity Fabric.
In particular, telemetric information about the state of individual cores, their temperature and consumption is collected via the same bus, and voltages and frequencies are controlled through it. In fact, Infinity Fabric can also be considered as a component of AMD's proprietary SenseMI technology.
AMD SenseMI Technology
An important part of Ryzen processors is a distributed network of current, voltage, consumption and temperature sensors, which allows you to accurately monitor the state of the processor. This telemetry data is collected on the Infinity Fabric bus every millisecond, allowing flexible control of the processor die while maintaining high responsiveness. SenseMI technology acts as an intelligent add-on to this mechanism. First, it manages the processor on the Infinity Fabric bus in a way that optimizes its instantaneous power and performance characteristics. Secondly, it also includes some functionality for prefetching and branch prediction. In general, SenseMI technology can be considered as a decomposition of several algorithms for different purposes.Mechanism Pure Power is responsible for saving energy and allows you to reduce the frequency and voltage of those processor units (or even cores), on whose contribution to the final speed of solving the problem nothing depends. In other words, thanks to Pure Power, the processor becomes more economical without any loss in performance.
Mechanism Precision Boost solves the opposite problem of Pure Power. Using telemetry data collected via the Infinity Fabric bus, it can increase the frequency of individual processor cores in small steps of 25 MHz, if this does not lead to the processor exceeding the established limits for temperature and consumption. In other words, Precision Boost is a flexible adjustment of the processor frequency to current conditions, similar to how modern video cards operate.
Technology Extended Frequency Range (XFR) is a mechanism that attracts the attention of enthusiasts for automatically overclocking a processor, depending on the parameters of its cooling system. XFR is implemented only in processors that have the ending X in their names. In them, subject to a number of conditions, it can further increase clock frequency beyond the limits set within Precision Boost. In most cases, XFR is activated if the temperatures of the processor cores are far from the limit values, however, in addition to absolute temperature values, XFR also focuses on their derivatives.
Neural Net Prediction– another facet of SenseMI technology. It means that the Zen architecture has a real real-time learning neural network that predicts how an application will behave in the near future. Such forecasting makes sense in order to proactively prepare instructions for execution and the data needed for them.
And the last part of SmartMI is the mechanism Smart Prefetch. It prefetches the necessary data in the processor's L1 and L2 caches based on information about how the application has worked up to that point. This eliminates possible processor downtime that may occur due to untimely loading of data.
Bottom line, there is no doubt that the Zen microarchitecture represents a giant step forward over Bulldozer. And the point is not only that the new processors use a modern technical process and a traditional x86 design with full-fledged wide cores without shared blocks and with support for multi-threading (SMT). A lot of other improvements have also been made, thanks to which the number of instructions executed by one core per clock cycle has increased by more than one and a half times. This is supported by improved branch prediction, the appearance of a micro-operation cache, the ability to send up to six micro-operations per clock cycle for execution (versus four), a 60% increase in scheduler buffers, a two-fold increase in the rate of completion and retirement of micro-operations, a one-and-a-half-fold increase in the depth of data loading and unloading queues, the ability to perform up to four floating-point operations per clock cycle (versus three), a multiple increase in the throughput of all caches and an increase in the size of the L1 cache, improvements at the level of data prefetching, and a lot of other things.
Test processor: AMD Ryzen 7 1700X
Today, March 2, 2017, AMD begins sales of the first batch of its fundamentally new Ryzen processors. And this is a truly historical event: there have not been products on the processor market for a very long time that would carry such a burden of expectations. It's no joke - AMD is going to compete with older Intel processors for high-performance desktops, but at the same time almost halve the price level.During the first phase of Ryzen's market launch, AMD is going to rely on its eight-core processors, classified as the Ryzen 7 family. These are the most expensive desktop media of the new Zen microarchitecture, costing from $330 to $500. But despite the relatively high price, the company expects almost rush demand for the new product and has seriously prepared for it. Product lots of Ryzen 7 are already in the warehouses of leading stores, and in total AMD has previously produced about a million processors.
In positioning new products, AMD adheres to several different principles than Intel. The company is clearly betting on greater mass appeal. At the same time, she sees the Ryzen 7 1800X as a twice cheaper alternative to the Core i7-6900K. The Ryzen 7 1700X is opposed not to an eight-core one, but to a similar-priced six-core one Core processor i7-6800K. The Ryzen 7 1700 is announced as a direct competitor to the quad-core Core i7-7700K. In other words, AMD's old tactics, when it tried to counter Intel's offerings with a superior number of cores at a lower price, are reflected in the new line. However, AMD's cores are now much more powerful than before, and the Ryzen 7 family looks very strong indeed.
To get acquainted with the new line of processors, we received from AMD a medium model, the Ryzen 7 1700X, which is interesting because it can be used to build configurations with a not too high cost - from 80 to 100 thousand rubles.
It is important to keep in mind that Ryzen processors are installed in a special new Socket AM4, which is now becoming the base for the entire range of AMD desktop processors. And this means that old motherboards are not suitable - we need new ones based on AMD X370, B350, etc. chipsets.
This is how the Ryzen 7 1700X is detected by the CPU-Z diagnostic utility.
Before us is a new 8-core processor from AMD, codenamed Summit Ridge and Zen microarchitecture, which stands out for its SMT support and the ability to execute 16 threads simultaneously, a 512 KB L2 cache per core and an L3 cache of two 8 MB parts.
The nominal frequency of the Ryzen 7 1700X is set to 3.4 GHz, but in most cases you can observe the operation of this processor at a frequency of 3.5 GHz - the Precision Boost technology has an effect. Moreover, with a low-threaded load, the frequency can increase to 3.8 GHz, and if you’re lucky, even up to 3.9 GHz due to XFR.
The supply voltage of our Ryzen 7 1700X under load ranged from 1.25-1.275 V. AMD says that the standard voltages for different Ryzen 7 can be set within a very wide range and typically range from 1.2 to 1.3625 V. This means that compared to 14nm Intel processors we will see higher voltages. That's why temperature regime The Ryzen 7 1700X at face value does not cause any particular concern. Under load, we observed heating up to 76-78 degrees according to the temperature sensor built into the core. At rest, temperatures are about 45 degrees.
Socket AM4 platform and new chipsets
As already mentioned, processors of the Ryzen family are focused on using a fundamentally new platform and a new Socket AM4 connector. This is primarily due to the fact that AMD had a need to introduce support for DDR4 memory, which has now become an industry standard. And at the same time, taking advantage of the moment, it was decided to redesign the entire platform, making the processors similar to SoC. In other words, an additional set of controllers was moved to the integrated north bridge of the processor, which made the new generation chipsets extremely simple devices.
As a result, it is not surprising that the new AM4 processor socket has an increased number of pins - there are now 1331 of them. This means that Ryzen is not compatible with any older motherboards. In addition, AMD has changed the requirements for the location of mounting holes for cooling systems on motherboards, and therefore Ryzen requires new coolers, or at least new mounts for old ones. Therefore, despite the fact that Ryzen at first glance is similar to its predecessors, has similar dimensions and external design, the entire ecosystem for them must be completely updated.
In Bulldozer processors, a memory controller was implemented in the processor chip. In the latest generations of APUs, the controller for the PCI Express graphics bus has also moved into the main chip. In Ryzen, additional PCI Express lanes, USB and SATA ports were added to the processor. In fact, AMD has now created a situation where the processor can operate without any additional chipset at all, which makes it possible to create extremely simple and compact motherboards.
However, it’s worth starting with the fact that the built-in memory controller in Ryzen processors is completely new. It is designed to work with dual-channel DDR4 SDRAM and supports only such memory. Backward Compatibility with DDR3 SDRAM is not provided. Officially, the Ryzen memory controller supports DDR4 modules with frequencies up to 2666 MHz, for which two or four slots can be provided on Socket AM4 motherboards. Memory with frequencies higher than DDR4-2666 with Ryzen can also be used, but the authors of the processor do not provide any guarantees in this case.
However, problems may arise when using high-speed memory modules in Socket AM4. The maximum DDR4 frequency that can be achieved in Ryzen without changing the base BCLK frequency is only 3200 MHz. Moreover, operation of DDR4-2933 or DDR4-3200 memory is possible only if a pair of modules is used. In other words, in terms of the frequency capabilities of the memory controller, Ryzen is much inferior to current Intel processors for the LGA 1151 platform, which freely conquer DDR4-4000 and higher modes. But for now there remains some hope that the situation can be corrected through new BIOS versions for motherboards.
In addition to an integrated memory controller with support for dual-channel DDR4 SDRAM, Ryzen provides:
16 PCI Express 3.0 lanes for a graphics card (if necessary, can be divided into two slots according to the formula 8x + 8x);
4 PCI Express 3.0 lanes for connection to the chipset or for other devices;
4 USB 3.0 ports;
4 PCI Express 3.0 lanes for NVMe storage (can be reconfigured into 2 PCI Express 3.0 lanes for NVMe storage and two SATA ports).
Thus, from just one Ryzen processor, a complete system-on-chip is obtained.
However, for typical desktop systems, the expansion capabilities available in the processor will most likely not be enough. Therefore, one of the logic sets - X370, B350 or A320, which will add some additional things to the specified list, can be connected to the processor via the PCI Express lines allocated for this purpose. And if there is no need for this, then it is possible to equip Ryzen with special simplified Mini-ITX chipsets X300 or A300, which do not consume PCI Express 3.0 processor lines, but also add almost nothing to the list of capabilities.
The bulk of the properties of the Socket AM4 platform are determined by the Ryzen processor. Chipsets in the new platform play a purely secondary role, and in fact, little depends on them in terms of the functionality of the platform.
Even the higher-end X370 chipset, which is likely to be used in most enthusiast motherboards, doesn't bring much: an additional two USB 3.1 ports, six USB 3.0 and USB 2.0 ports each, eight SATA ports, four of which can be converted two SATA Express interfaces, and eight additional slow PCI Express 2.0 lines. Plus, in the Socket AM4 platform, the use of one or another chipset either allows or prohibits overclocking, dividing graphic lines PCI Express 3.0 x16 and RAID modes for SATA ports. For example, in the same X370, as in the older chipset, overclocking, SLI or CrossfireX configurations, and RAID arrays of levels 0, 1 and 10 are allowed.
Along with the X370, advanced users may also be interested in the simpler B350 logic set. It still allows processor overclocking and RAID arrays, and the main difference from the older version concerns the impossibility of dividing the processor graphics bus into two slots. In addition, some of the USB 3.0 and SATA ports fell under the knife, of which there are two and six left in the chipset, respectively, plus the number of PCI Express 2.0 lanes was reduced to six.
Another interesting alternative is the X300, a chipset that is specifically designed for simple compact systems. It doesn’t add anything to the processor’s capabilities, but it allows the PCI Express 3.0 x16 graphics bus to be divided into two slots and allows overclocking of the processor.
We have summarized detailed information about what capabilities each chipset offers in combination with Ryzen in the following table.
Although the logic sets bear the AMD name, ASMedia, known for its various controllers, played a primary role in their development. It was thanks to her that AMD was able to be the first to bring to market chipsets that support USB 3.1 ports with a bandwidth of 10 Gbps. However, innate support Type-C connectors while in AMD chipsets No. In order for a convenient symmetrical USB connector, motherboard manufacturers will have to fork out for an additional chip driver.
Thanks to support for USB 3.1, chipsets for the Socket AM4 platform look modern, but you still shouldn’t delude yourself too much about their capabilities. While Intel's 200 series chipsets can provide up to 30 high-speed ports (PCIe 3.0, SATA and USB 3.0), even the older AMD X370 has half as many such ports. This is partially compensated by the capabilities of the northbridge built into the processor, but nevertheless, the Intel platform allows you to create more flexible configurations with greater possibilities for connecting additional devices.
For testing, we received maternal ASUS board Crosshair IV Hero.
This motherboard is based on the older AMD X370 chipset and uses its potential to the maximum. The board supports the division of the PCI Express 3.0 graphics bus into two slots and configurations built using SLI and CrossfireX technologies. Both graphics slots on this board are reinforced with SafeSlot metal frames and are widely spaced to accommodate massive and powerful GPUs.
The board supports overclocking, and its overclocking settings are made so that operating the processor at higher frequencies does not cause problems. To cool system components, Fan Xpert technology is provided, which allows you to control all five fans that are connected to the board. Like the latest ROG series boards for LGA 1151, ASUS Crosshair IV Hero has dedicated connectors for connecting a liquid cooling system pump, as well as temperature and coolant flow rate sensors. There is also a special connector for high-power fans.
An important feature of Ryzen-based systems is that the M.2 slot for NVMe drives connects directly to PCI Express 3.0 processor lines. This is exactly what was done on the Crosshair IV Hero. There are no speed restrictions - M.2 has four necessary PCIe lanes. At the same time, the M.2 slot itself is located further away from the processor and video cards - where it will be easier for it to organize adequate cooling.
The board is equipped with the now fashionable RGB illumination, which is controlled via the ASUS Aura RGB application. You can also connect additional LED strips to the Crosshair IV Hero.
Integrated sound card based on an exclusive codec latest generation S1220, which provides a signal-to-noise ratio of 113 dB. This codec works in conjunction with the premium ESS Saber DAC, which in total allows you to get sound quality comparable to that provided by inexpensive discrete ones. sound cards. In addition, the Sonic Studio III program is included with the audio path, allowing you to easily manage audio streams. For example, it can be used to send sounds from a game to headphones, music to speakers, and sound from a video to a TV.
In short, then ASUS specifications Crosshair IV Hero looks like this:
The gigabit network on the board is represented by the usual Intel controller, which is equipped with the GameFirst program for prioritizing network traffic. In addition, the board has an additional M.2 slot into which you can install a WiFi controller.
The rear panel of the board is densely filled with ports, plus the Clear CMOS and BIOS Flashback hardware buttons have been moved to it. But the main area is occupied by numerous USB ports, among which there is a 10 Gbps USB 3.1 port in Type-A and Type-C variants. By the way, the board also provides a pin for a USB 3.1 port, which is located on the front panel of the case.
The recommended price for ASUS Crosshair IV Hero is $255.
How we tested
Testing of the AMD Ryzen 7 1700X processor was carried out in full accordance with the manufacturer’s precepts: AMD’s flagship product was opposed to the entire current line of Core i7 processors. In addition, we did not forget to include the senior processor of the AMD FX line in the tests.Ultimately, full list The components involved in the test systems received the following form:
Processors:
AMD Ryzen 7 1700X (Summit Ridge, 8 cores + SMT, 3.4-3.8 GHz, 16 MB L3);
AMD FX-9590 (Vishera, 8 cores, 4.7-5.0 GHz, 8 MB L3);
Intel Core i7-7700K ( Kaby Lake, 4 cores + HT, 4.2-4.5 GHz, 8 MB L3);
Intel Core i5-7600K (Kaby Lake, 4 cores, 3.8-4.2 GHz, 8 MB L3);
Intel Core i7-6900K (Broadwell-E, 8 cores + HT, 3.2-4.0 GHz, 20 MB L3);
Intel Core i7-6800K (Broadwell-E, 6 cores + HT, 3.4-3.8 GHz, 15 MB L3).
CPU cooler: Noctua NH-U14S.
Motherboards:
ASUS Crosshair IV Hero (Socket AM4, AMD X370);
ASUS 970 PRO Gaming/Aura (Socket AM3+, AMD 970 + SB950);
ASUS Maximus IX Hero (LGA1151, Intel Z270);
ASUS X99-Deluxe (LGA2011-v3, Intel X99).
Memory:
2 × 8 GB DDR4-3000 SDRAM, 15-17-17-35 (Corsair Vengeance LPX CMK16GX4M2A3000C15).
4 × 4 GB DDR4-3000 SDRAM, 15-17-17-35 (G.Skill F4-3000C15Q-16GRR).
2 × 8 GB DDR3-2133 SDRAM, 9-11-11-31 (G.Skill F3-2133C9D-16GTX).
Video card: NVIDIA GeForce GTX 1080 (8 GB/256-bit GDDR5X, 1607-1733/10000 MHz).
Disk subsystem: Kingston HyperX Savage 480 GB (SHSS37A/480G).
Power supply: Corsair RM850i (80 Plus Gold, 850 W).
Testing was performed in the operating room Microsoft system Windows 10 Enterprise Build 14393 using the following driver set:
AMD Chipset Driver Crimson ReLive Edition 17.2.1;
Intel Chipset Driver 10.1.1.38;
Intel Management Engine Interface Driver 11.6.0.1030;
Intel Turbo Boost Max Technology 3.0 1.0.0.1029;
NVIDIA GeForce 378.66 Driver.
Performance
Comprehensive PerformanceTo evaluate the performance of processors in common tasks, we used the BAPCo SYSmark 2014 SE test package, which simulates user work in real common modern office programs and applications for creating and processing digital content. Latest versions this benchmark operates on four scenarios: Office Productivity (office work: text preparation, spreadsheet processing, work with by email and visiting Internet sites), Media Creation (working on multimedia content - creating a commercial using pre-filmed digital images and video), Data/Financial Analysis (processing an archive of financial data, their statistical analysis and forecasting investments based on a certain model) and Responsiveness (analysis of the system’s responsiveness when launching applications, opening files, working with an Internet browser with big amount open tabs, multitasking, copying files, batch operations with photos, encrypting and archiving files and installing programs).
AMD contrasts the Ryzen 7 1700X with the six-core Core i7-6800K processor, but as we can see, according to the integrated indicator in SYSmark 2014 SE, the new AMD product is still inferior to it, demonstrating the performance level of the Core i5. The problem is that most commonly used applications remain single-threaded, and with such a load, Ryzen is still weaker than Intel architectures, although not by much. A clear illustration of this can be seen in the results of the Office Productivity script. In complex multi-threaded workloads, especially those of a counting nature, the Ryzen 7 1700X’s performance is fine. Thus, in the Data/Financial Analysis subtest, the new Ryzen 7 1700X not only outperforms the six-core Core i7-6800K, but also turns out to be stronger than Intel’s eight-core Core i7-6900K.
To evaluate complex performance in gaming 3D, we used the Futuremark 3DMark Professional Edition 2.2.3509 test, in which we used the Time Spy 1.0 scene.
This benchmark is well optimized for multi-threading, so the Ryzen 7 1700X demonstrates very good speed in it. The Zen microarchitecture allowed AMD to make a full-fledged eight-core processor, and its performance is closer to the Core i7-6900K than to its direct competitor, the Core i7-6800K.
Tests in applications
The task that responds most sensitively to increasing processor parallelism is traditionally final rendering in 3D design and modeling packages. We tested rendering speed in two popular applications: Autodesk 3ds max 2017, where we measured the time spent rendering at a resolution of 1920 × 1080 using the mental ray renderer of a standard Hummer scene; and in Blender 2.78a where the duration of building the final model from Blender Cycles Benchmark rev4 was checked.
The Ryzen 7 1700X fully delivers on its promises and delivers rendering performance that was previously only possible with eight-core Intel processors. However, it should be recalled that the Ryzen 7 1700X costs about two and a half times cheaper than the Core i7-6900K.
The next test task is image processing. Adobe Lightroom 6.8 is used here and Adobe Photoshop CC 2017. In the first case, performance is tested when batch processing a series of images in RAW format. The test scenario involves post-processing and exporting to JPEG at 1920 × 1080 resolution and maximum quality of two hundred 12-megapixel RAW images taken with a Nikon D300 digital camera. In the second - performance when processing individual graphic images. To do this, we measure the average execution time of a test script that is a creative reworking of the Retouch Artists Photoshop Speed Test, which involves typical processing of four 24-megapixel images taken with a digital camera.
Adobe applications for photographers - with special features. In Photoshop, many filters and operations are still performed in a single thread. Lightroom began to actively use AVX2 instructions. Both are bad for the Zen microarchitecture, so in both test tasks the Ryzen 7 1700X processor loses even Quad Core i5, not to mention higher-end Intel processors.
But video processing, like rendering, is considered a task whose performance scales well with increasing processor parallelism. Here we used four tasks for testing. Adobe After Effects CC 2017 – testing rendering speed using ray tracing. The time spent by the system on rendering a pre-prepared video at 1920 × 1080@30fps is measured. Adobe Premiere Pro CC 2017 - performance testing for non-linear video editing. The time for rendering a Blu-Ray project containing HDV 1080p25 video with various effects applied is measured. x264 r2744 - testing the speed of video transcoding into H.264/AVC format. To evaluate performance, we use an original 1080p@50FPS AVC video file with a bitrate of about 30 Mbps. And x265 2.2+17 8bpp - testing the speed of video transcoding into the promising H.265/HEVC format. To evaluate performance, the same video file is used as in the x264 encoder transcoding speed test.
When working with video, as well as in final rendering, the Ryzen 7 1700X is very good. It can really compete with the thousand-dollar Core i7-6900K, which makes AMD's new product an ideal choice for users who create multimedia content.
To measure processor performance when compressing information, we chose two archivers: 7-zip 16.04 and WinRAR 5.40. In both cases, the time spent on compression with the maximum degree of compression of the directory was measured with various files with a total volume of 1.7 GB.
For fast operation of archivers, good throughput and low latency of the memory subsystem are important. The memory controller of Ryzen processors turned out to be extremely unsuccessful, so in these tests the Ryzen 7 1700X can only be compared with Intel quad-core processors.
Browser performance Microsoft Edge was tested in a specialized test, WebXPRT 2015, which implements algorithms actually used in Internet applications in HTML5 and JavaScript.
The task is single-threaded, but the Ryzen 7 1700X holds up at a good level, second only to Intel processors based on the Kaby Lake microarchitecture.
In conclusion, we checked the speed of cryptographic algorithms in the VeraCrypt 1.19 utility. Here, the benchmark built into the program was used, using Serpent-Twofish-AES triple encryption.
The task is single-threaded, plus the implementation of the AES instruction set in Zen is very efficient. The result is not long in coming: Ryzen 7 1700X is in first place.
Gaming Performance
Until recently, the performance of platforms equipped with modern processors in the vast majority of current games was determined by the capabilities of the graphics subsystem. However, the rapid growth in productivity over the past few years gaming video cards led to the fact that now performance is often limited not so much by the video card, but central processor. And if earlier, in order to understand the gaming potential of a particular CPU, we had to use reduced resolutions, then with modern video cards this is not at all necessary.
To complete our processor test system, NVIDIA provided us with its latest GeForce GTX 1080 accelerator, which, thanks to its unprecedented high power, is well suited for 4K resolutions, virtual reality, and even more so for FullHD. As a result, we were able to abandon gaming tests in a resolution of 1280 × 800, which were often not understood by our readers. Now the dependence of the frame rate on CPU power can be perfectly traced in absolutely real, and not artificially created, conditions: in FullHD resolution 1920 × 1080 and with maximum settings image quality. This is the approach we adopted.
Games don't give much reason for optimism regarding Ryzen. No, of course, these are not FX series processors, whose gaming performance has already become a reason for ridicule. The Ryzen 7 1700X delivers a level of gaming performance that is more than acceptable at the present stage, and it can certainly handle GeForce GTX 1080-class video cards without question. But if you look at relative performance indicators, it turns out that any current Intel Core i7 and even Core i5 processors have higher gaming potential - with high quality This is visible in graphics even in the most common FullHD resolution. The reasons for this state of affairs are well understood: the slow Ryzen memory controller and the weaker operating speed of the FPU part than that of Intel processors.
However, it must be emphasized once again that at the moment the power of the Ryzen 7 1700X is perfectly sufficient to provide high frame rates in games. And therefore it is considered insufficiently productive gaming CPU still shouldn't. In addition, the new AMD product has eight full-fledged cores, which can be a good help in new gaming projects that, although timidly, are still moving towards the full use of multi-threading and the transition to DirectX 12.
Energy consumption
The power consumption situation is another intriguing section of today's testing. AMD has moved its processors to a modern 14nm process technology and optimized the architecture with a clear focus on energy efficiency. As a result, the company now says eight-core Ryzens fit into a 95-watt TDP. That is, they should be noticeably more economical than Intel LGA 2011-3 processors with a typical heat dissipation of 140 W. Has the situation with real power consumption become the place where the Ryzen 7 1700X can win an unconditional victory over its competitor? Let's check.The new one we used in the test system digital block Corsair RM850i power supply allows you to control the consumed and output electrical power, which is what we use for measurements. The graph below shows the total system consumption (without monitor), measured “after” the power supply and representing the sum of the power consumption of all components involved in the system. The efficiency of the power supply itself is not taken into account in this case.
When idle, the Socket AM4 platform really looks very economical. And this is not surprising, Ryzen uses advanced energy-saving technologies, and does not have any special energy appetites and the accompanying logic sets.
But when rendering in Blender, the consumption situation looks a little different than expected. Under load, a system with a Ryzen 7 1700X requires about the same amount of energy as a platform based on a Core i7-6900K. And this raises doubts that the Ryzen 7 really fits into the 95-watt thermal package.
And here’s what the consumption situation looks like at the maximum possible load: in the Prime 28.10 utility, which actively uses extremely energy-intensive FMA and AVX2 instructions.
In terms of maximum consumption, the Ryzen 7 1700X still manages to lag slightly behind the Core i7-6900K. We are, of course, not talking about the 30 percent difference that is mentioned in the specifications, but about a difference at the level of just a few watts. In theory, the Ryzen 7 1700X should have been closer to the Core i7-7700K, whose thermal package is set at 91 W, but in practice, AMD’s offer is noticeably more power-hungry.
Overclocking
Unfortunately, Ryzen is chasing poorly. It is obvious that the nominal frequencies of these processors are raised to the limit at the factory. Therefore, one cannot count on the fact that productivity can be further increased with simple manipulations.The stable maximum that we managed to achieve with our copy of the Ryzen 7 1700X was only 3.85 GHz, that is, we were able to go beyond the turbo mode only a little. The processor no longer took on a higher frequency.
And even then, in order for the system to pass stability testing in Prime 95 10/28, the processor supply voltage had to be raised more than seriously - up to 1.5 V. The fact is that long-term operation of a 14-nm chip at such a voltage will not lead to to the degradation of the semiconductor crystal, there are well-founded doubts.
In addition, the temperature regime turned out to be not very favorable with such a seemingly insignificant overclocking. Despite the fact that Ryzen has solder under the lid and not paste, the thermal sensor built into the processor chip recorded heating up to 99 degrees.
conclusions
We all really hoped for it, and it happened: AMD did it. The new Ryzen processors are radically different from Bulldozer. The microarchitecture in them has been completely updated, and now Ryzen 7 is a product high level. As promised, single-threaded performance in the new product has increased by about one and a half times, and power consumption has decreased by about the same. As a result, AMD has a high-performance eight-core processor that can really be put on the same level as Intel's offerings for the LGA 2011-3 platform. In addition, AMD seems to have very ambitious plans in light of its return to the market, since at the same time it is trying to break established pricing and begin to offer high-quality eight-core processors at unprecedentedly low prices.As a result, the new AMD platform could be a very attractive solution for those users who require high multi-threaded performance. As our extensive tests have shown, top scores Ryzen 7 shows when working on digital content - during rendering and video processing. This means that professionals and hobbyists who choose configurations for work rather than play should seriously consider choosing Ryzen 7 processors for themselves. However, this recommendation does not apply to photographers: with graphic editors The new AMD microarchitecture is not performing well.
As for the more widespread applications of computers - games, then for them Ryzen is far from best choice. There are two weak points in the design of new AMD processors: the memory controller and the relatively weak FPU unit. Both are very important in gaming tasks. Therefore, eight-core AMD processors in them provide only Core i5-level performance. Of course, this is by no means a death sentence, because this speed is generally quite sufficient for modern graphics cards.
And yet, based on the results of the review, we can say that Ryzen 7 is a clear success for AMD. The company is returning to the upper price segments, and nothing more is needed yet. Let's hope that the company's engineers will now be able to adhere to the schedule they themselves set and will annually release improved versions of Zen, in which all the bottlenecks of this microarchitecture will be gradually corrected.
Today is precisely the case when thousands of words could be written in the introductory part of the article. Of course, Ryzen is coming out - the most promising x86 processor in the last five years, which is also of great importance for the path the personal computer industry will take in the near future. However, you are probably expecting us to talk at length about how much we expect the new product and how good it would be if full-fledged competition returned to the processor market. Therefore, we will not postpone the most interesting things until later, but will immediately move on to the technical details, and then to the tests.
The way the Ryzen 7 1800X overclocks (or rather, does not overclock) I really want to attribute to the dampness of the platform. It was with great difficulty that we managed to achieve stable operation of this processor at frequencies even slightly higher than the nominal values. In overclocking, progress in frequency is very sluggish, and further raising the V CORE voltage, taking into account the fact that it already exceeds 1.4 V at nominal value, and even fluctuates greatly over a wide range, is somewhat scary.
The stable maximum that was achieved was only 4.0 GHz. The processor no longer took on a higher frequency. The system booted up to a frequency of 4.25 GHz, but, unfortunately, there was no talk of any operation without crashes and freezes. To test, we used the Prime 95 10/28 utility, and it managed to crash the system in just a few minutes, even if the frequency was selected at 4.05 GHz.
However, the operation of the Ryzen 7 1800X at 4.0 GHz caused some concern. Firstly, in order for the system to pass stability tests, the CPU supply voltage had to be raised to 1.55 V. There are well-founded doubts that long-term operation of a 14-nm chip at such a voltage will not lead to degradation of the semiconductor crystal. Moreover, with each reboot, the motherboard complained about the processor voltage being dangerously high.
Secondly, the temperature of the CPU running at such an overclock, as reported by the built-in sensor, went off scale over 100 degrees, despite the fact that a powerful Noctua NH-U14S cooler was used for cooling in our experiments. This did not cause any throttling, but temperatures of about 105 degrees are not very similar to safe heating. Especially if you take into account the fact that Ryzen’s processor cover is soldered to the semiconductor chip, and not mounted on paste, like the competitor’s LGA1151 processors.
As a result, overclocking the Ryzen 7 1800X was able to increase the frequency by only 8-10 percent relative to the nominal value. Such a modest result did not allow us to go beyond the boundaries of the turbo mode frequencies, but the safety of even such a modest increase in frequency in the context of constant use of the system is a big question. All this leads to the disappointing conclusion that the overclocking potential of new AMD processors is frankly low, and Ryzen loses here to its competitor’s processors. For example, the same Core i7-6900K exceeds its nominal frequency by 20-25 percent and, when air-cooled, is capable of reaching the 4.2 GHz bar, which is beyond the capabilities of the Ryzen 7 1800X.
However, there is still little hope that the cause of such overclocking suffering is the “dampness” of the platform. For example, AMD itself promised something completely different in terms of overclocking. According to statements from company representatives, its new 14-nm processors should be able to overclock with air cooling to 4.2-4.3 GHz with voltages of about 1.45 V. Our experience so far categorically refutes these promises, but there is some hope for an improvement in the situation -it still remains. Therefore, we will return to the topic of processor overclocking in our future articles.
Experiments on overclocking the Ryzen memory subsystem also failed to become a source of optimism. The maximum DDR4 mode, which allows you to set the Ryzen 7 memory controller without increasing the BCLK frequency, is DDR4-3200. But even in DDR4-2933 mode, not all modules work with this processor. For example, the 2 x 8 GB DDR4-3200 Corsair Vengeance LPX CMK16GX4M2B3200C16 kit, which we use in testing Intel systems, ran on a Socket AM4 system with a Ryzen 7 1800X only in DDR4-2400 mode.
In return, AMD provided us with another, similar kit of similar volume, Corsair Vengeance LPX CMK16GX4M2B3000C15. It is designed for DDR4-3000 frequency, and with it we were able to run all tests in DDR4-2933 mode. However, any attempts to make it run at higher speeds failed. In other words, for now the situation looks as if in order to run the Ryzen memory subsystem on high frequencies, some special “selected” modules are needed. However, here too there remains hope that optimization will help over time. Motherboard BIOS plat.
In addition to the above, mention should be made of the special AMD Ryzen Master utility, which the company's engineers released to manage overclocking of new processors from operating system. However, unfortunately, it is not able to improve the results of overclocking and only adds some convenience to this process, allowing in some cases to do without constant reboots and tedious selection of settings in the BIOS environment.
In addition, the AMD Ryzen Master's feature set is somewhat limited. It only allows you to change the frequency of processor cores, voltage V CORE , as well as memory frequency and timings. Moreover, often after changing the parameters, a system reboot is still required for them to take effect. In addition, while the utility is in beta status, it distorts a number of parameters and does not display a number of them at all. So it will be possible to fully use it only after all the shortcomings and shortcomings are corrected by the developers.