In recent months, AMD has released a lot of information about the new Zen architecture and related chipsets. Time for an update where we set up for an overview regarding the architecture which AMD calls the most competitive in ten years.
Some time back we spent twice attention to the long-awaited Zen architecture from AMD, but in the meantime there have been several presentations in which the chip maker has released much more information. So we know now reported much more about the internal workings of Zen and there are several chipsets.
Zen is a new architecture designed to make competitive again AMD processors of Intel. AMD has been directed toward the improvement of the IPC and the efficiency relative to the now elderly Bulldozer-architecture, where the focus on higher clock rates and more cores at the expense of IPC and single-threaded performance led to an even further deterioration of AMD’s position on the processor market. In addition to the new architecture, AMD also makes a step from 32nm to 14nm process, which should provide a significant improvement in efficiency.
Zen is a comprehensive architecture, similar to Skylake and other Intel architectures. It is intended that there will come to Zen based processors for laptops, desktops and servers. Initially, we can expect the 8-core high-end desktop (HEDT) processors, code-named Summit Ridge. Later would appear too Summit Ridges with a six and four cores.
There is a lot of optimism about Zen, but also much skepticism. AMD has a lot of promises made in the last ten years who could not live up to the company later, especially in the Bulldozer debacle. We therefore look forward to the introduction of Zen, which will take place in the first quarter of 2017. Perhaps the processor market is again more exciting and interesting!
More detailed information on specific aspects of Zen can be found in several other preview articles we’ve published previously. A discussion of the architecture can be found here, and more information about AMD’s allegations concerning the IPC improvement in this article. This article aims all the information available about Zen to bring together in one place.
AMD has not yet released specific details about the Summit Ridge HEDT processors, but the laws are already rumors about this. Of course there is the ‘base’ Zen eight cores and AMD itself has indicated that there will at least be an 8-core variant of Summit Ridge on the market. There are also rumors that six-core and quad-core versions will come true – logical, because the company can not all chips are still selling perfect. The TDP of the 8-core would be 95 watts, compared with 65 watts for the 4-core variant. In both cases a giant step in efficiency for AMD, but an even larger step in the case of the 8-core – since the HEDT processors from Intel devoid al 140 watts. Assuming that what we will see will match the specified performance levels, it sounds almost too good to be true. It is worth mentioning that AMD has expressly refused to confirm this TDP values.
The different SKUs would go through life as SR3, SR5 and SR7 – similar to the i3 i5 i7 classification used by Intel. Since the SR stands for Summit Ridge, this would mean that the processors each generation will have to be renamed, so it is not clear to the designation which processor after another occurs. In contrast to Intel will there always be a large discrepancy between the processors, while there i7 exists with two cores to ten cores. The clock rate of the 8-core flagship model will be rumored 3.2GHz with a turbo 3.5 GHz. In itself impressive, but can get by the relatively low turbo single threaded performance will be less than Broadwell-E, which thanks to Turbo Boost 3.0 clock speeds of 3.8 and 4 GHz.
Unlike Intel’s mainstream processors Summit Ridge is not equipped with an integrated GPU. Useful for people who still go places a graphics card in the system, such as gamers. This means that there is less die-area is wasted on something that is not used, and therefore there is more available, for example, additional cores and cache. To give just one example, 40% of the size of Skylake quad-core processors used in the IGP – the same as for the four cores (the rest of which is different controllers that are integrated into the CPU).
AMD might have some trouble getting the clock speed of the quad-core models at an acceptable point, the TDP limit. Fortunately, the company still has some room for maneuver: even though Intel does hefty quad core models with a TDP of 65 watts, has a TDP of 95 watts certainly not out of the question for quad cores – especially when they are put on the market at a favorable price . Bulldozer had an exorbitant consumption, but it is not necessary to shoot in the other direction at the expense of performance improvements.
The prices have not been announced, although it is obvious that they will be competitive compared to what Intel currently asking for its processors. An alleged price list of a Chinese partner of AMD indicates that the 8-core processors depends on the binning should cost between 350 and 500 dollars, while the 6-core and 4-core processors will cost 250 and 150 dollars respectively. The suggested retail prices for the i7 6700K, 6800k and 6900k are respectively 339, 434 and $ 1089.
Shortly after the Summit Ridge processors will still hatch models with an integrated GPU and Zen cores, presumably with a TDP of 45 watts and 65, called Raven Ridge. The structure of the Zen modules makes it possible to easily design processors for the lucrative server market, because there are so several blocks with eight cores on one that can be installed (which you read about later). Both servers and laptops is a prerequisite for success, however, that it is better to energy efficiency, as this is still more important than for desktops in those markets.
Improvements in the architecture and the smaller process would Zen much better performance per clock (IPC) must have Bulldozer. AMD has indicated that Zen must have a 40% better IPC, although “more than 40%” is sometimes also asserted during presentations. This is not relative to the current high-end FX processors, but compared with Excavator cores, which we encounter in Carizzo and Bristol Ridge processors.
AMD has further indicated that the performance should be comparable with Broadwell-E processor. With an AMD demo of the Blender program proved an equally clocked Zen even slightly faster than Broadwell-E. This of course does not mean you can assume that Zen is as fast as Broadwell-E: This type of presentation is notorious for cherry-picking of the benchmarks. Even with Bulldozer we saw similar promises. Nevertheless, AMD claims to be aware of his reputation for overpromising and under delivering and would heavily hyping products should belong to the past. This guarantees nothing, but it is certainly a good sign.
A while back we IPC the only desktop Carizzo (the Athlon X4 845) and compared Broadwell-E, and revealed that AMD must take a bigger step than 40% in order to arrive at the same IPC. Broadwell-E is the same clock rate is approximately 70% faster than Carizzo. Whether it is 40%, 70% or something in between, even 40% is already a huge step forward in terms processor and this will lead to a lot more competition in the processor market, which in recent years was unfortunately little evidence of it.
All parts of Zen are fully designed from scratch again. The architecture is baked on the 14nm process of GlobalFoundries, which AMD is a big step from the 32nm which it currently uses for its HEDT processors. On the basis of the alleged size would you might suspect that this is equivalent to the Intel 14nm process, but that is not the case. Intel process has smaller effective size than other semiconductor manufacturers being able to produce.
One of the main, if not the most important change in Zen is getting off the Bulldozer modules. These were two “cores”, each with only a single FPU (floating point unit) and a shared floating point scheduler. AMD assumed this would satisfy, because floating point (or decimal) numbers are used less than integers (whole numbers). Another assumption was that multithreading was about to break, and that more slower cores could compete with a smaller number faster cores. This turned out not to be so.
At Zen AMD again fit to a traditional design, similar to the Phenom (K10) processors who came before Bulldozer, albeit a much wider and more sophisticated. AMD processors will thus go a lot more similar to those of Intel. Where each Bulldozer core had to share his floating point scheduler with the other core in the module, each core has now headed his own floating point scheduler. The number of floating point units has increased significantly: in each Bulldozer core was effectively a single unit, but now there are four (two for two for addition and multiplication).
Furthermore, the number of units for integers is now larger pieces. Where bulldozer had four integer execution units, two calculations (ALUs) and two memory operations (AGU’s), Zen has four ALUs and two AGU – one AGU less than Skylake.
There is also another flat major improvement: the width of the pipeline. This is 50% wider than Bulldozer and can handle 75% more instructions. The number of items that fit in the “queue” (queue) has also increased. There may at Zen 84 integer instructions and 96 floating point instructions are scheduled, compared with 48 and 60 by bulldozer. The queue for retrieving data from cache or memory has been expanded, but at this point AMD gives little detail.
In design is Zen certainly not identical to the architecture used by Intel. For example, Intel sit at all cores side by side, with the shared L3 cache in the middle. Thanks to the ring bus, however, all the cores have access to any part of the cache. This is true for the 4-core processors, but also for the 10-core 6950X. Zen processors are a little bit different: a zen-processor consists of one or more so-called CPU Complex’en (abbreviated CCX). Here you have four cores and 8MB of L3 cache, which is very similar in format to the quad-core processor from Intel. In an 8-core Zen, there are two of them placed side by side, so that it looks a lot different than the Intel 8-core processors.
All cores can access the cache within the same CCX with the same average speed, but this does not apply to the cores in another CCX. It remains to be seen what impact the performance of this design, especially with models sold as quad cores by several broken cores.
In addition to leaving the bulldozers modules is introduced at Zen Simultaneous Multi Threading (SMT), which each core has two threads. SMT is intended to make multithreading efficient by as many units to engage inside the processor, where mostly very CPU-intensive applications such as encoders benefit.
Meanwhile, there is little known about how AMD SMT will work properly. Thus expanded branch predictor, which should ensure that branches are predicted well in the code and that therefore more instructions can be executed in parallel. In addition, Zen gets a micro-op cache, which should ensure that the instructions that the processor executes accelerated when the code contains many identical instructions.
An open question is how well AMD’s implementation of SMT will be. Intel’s implementation of SMT called HyperThreading, for example, it took some time before it in virtually any program or game led to poorer performance. The possibility therefore exists that AMD a similar learning curve will know where AMD’s version of SMT does not immediately work as well as HyperThreading after many years of optimization. In any case, it is a step in the right direction, compared with a profit of core Bulldozer.
Also in terms of processor instructions Zen is well up to date. Naturally all SSE instruction sets are supported. He offers AVX and AVX2, which the processor can handle 128-bit and 256-bit numbers. Intel’s AES instructions and SHA1 / SHA256 sets are implemented, that encrypting data and creating hashes is much faster, respectively. Further, the instruction RDSEED also useful for encryption, the pseudo-random number generator that Intel introduced with Ivy Bridge.
The processor cache immediately started to kick. All three levels of cache are faster compared to Bulldozer. The L1 and L2 caches are up to two times faster, while the total bandwidth of L3 cache, according to AMD even five times higher.
The size of the caches is also changed. Compared to its predecessor, Zen twice the L3 cache, with 16MB for an 8-core model, compared to 8MB for Bulldozer. Interestingly enough, the amount of L2 cache or halved: from 4MB to 8MB. Bulldozer had 48 kB total L1 cache, which consisted of 32 KB instruction cache (actually 64K per module) and 16 KB data cache. When Zen is 64 KB instruction cache and 32 KB data cache per core, which allows the instruction cache is twice as large as what Intel currently offers.
Furthermore, the L1 cache now from the write-back type, so the data is not written automatically to the L2 and L3 cache needs to be written to it. This should have a positive effect on the rate and efficiency.
In addition, the pre-fetchers are greatly improved. This is the component that predicts which short-term data is needed for executing instructions, and then this already retrieves from the relatively slow DRAM. With better pre-fetching the CPU cores will be less time out of work while waiting for data.
Not only the capabilities of the processor are important, but also that of the chipset. Since AMD has now also the necessary information released. This same chipsets will be clear that you can combine with the Bristol Ridge processors (based on Excavator cores), a newly launched family of 28nm APUs. The AM4 socket will be a “universal” socket where all processors from AMD can be placed.
There will be three different chipsets that support Zen X370 namely, B350 and A320. There is a still further AM4-series chipsets, namely X / B / A300, but these are intended for systems on a chip, and therefore do not provide their own functionality. It is to be seen whether Zen will even work here.
The B350 and A320 we already know what they will provide, and that is unfortunately what scanty. The B350 chipset provides up to six times USB 2.0 and two USB 3.1 and 3.0. He also supports two SATA ports and a SATA Express port. The number of available PCI Express lanes of the chipset is small: six, and are also PCI Express 2.0 lanes with a lot less bandwidth than 3.0. A320 is largely identical, but offers a USB 3.1 port less and has only four PCI Express 2.0 lanes. SLI / CrossFire (ie the ability to split processor lanes) and overclocking are not possible in these two chipsets, therefore you should have the X370. AMD has said nothing officially about this, but according to rumors he would be coupled with a Zen-processor offering the ability to use two graphics cards with 16 lanes.
There are rumors that overclocking will indeed be possible with B350, but this was never officially confirmed by AMD, and slides even suggest that this will not be the case.
It seems that the B350 and A320 are primarily intended for the Bristol Ridge APUs, which can drive itself extra SATA and USB ports and therefore need to rely less on the chipset for these opportunities. It is certainly not impossible that Zen also additional features will have on board – remember that AMD often was the first to integrate functionality into the CPU. Nevertheless, we consider it unlikely, because there are more important things that AMD has to keep working to make a success of Zen. The risk therefore exists that the X370 of a very different order than the other two, but the difference should be much greater than between Z170 and H170.
The story obviously does not end with Zen. AMD has plans for the successor of Zen called Zen + yet at this time. Zen + should be an optimized version of the Zen architecture, as the name already suggests. This will most likely not refresh or revolution, but simply an incremental and gradual improvement as we have seen many times from the company. Consider the step of Bulldozer to Piledriver, which is still a significant performance entailed.
Will reportedly Zen + baked at 7 nm process of GlobalFoundries. 10nm is about beaten by the semiconductor manufacturer, because it would not provide sufficient returns to justify the huge investments that require modern chip factories. In any case, AMD and GlobalFoundries recently decided that AMD is no longer obliged to take all his chips away at GlobalFoundries. If competitor TSMC therefore an interesting example develop 10nm process, if there are problems at GlobalFoundries (as so often in the past), then AMD has a way out.
Just as the 28nm to 14nm step you must moreover not imagine that here there will be the effects of a ‘double’ which shrink as the actual dimensions nowadays very much different from what state semiconductor manufacturers. We should be glad if the effects are as large as those of some who shrink.
Of Zen is a success or not, the performance difference between AMD and Intel will probably shrink in the coming years. Intel has most of the “low-hanging fruit” already picked, the relatively simple ways to be bring about performance improvements. This is not to this extent in the case of AMD processors. If you compare the current generation of AMD processors with Skylake, they have been an IPC which 50% lower. AMD is at a historic low, if you compare the performance with that of Intel.
This backlog to make up is a lot easier than coming up with breakthrough improvements (including AMD in the past, moreover, has contributed more than its share). Indeed, it is a familiar way: AMD now knows like the rest of the world Simultaneous Multithreading is a lot faster than the disastrous process of the Bulldozer modules. One way to restore the self-inflicted wounds AMD will be able to make a much more competitive.
Moreover, the focus of Intel is long gone on pure performance. The company considers the mid-range and high-end desktop market as safe, it introduced a change in Broadwell in the formula for acceptable architecture improvements. Previously, a change was made as it resulted in 1% performance and a maximum 1% higher power consumption, and thus a constant performance per watt. From Broadwell may 1% performance increase power consumption up to 0.5%. Architecture changes should now result in better performance per watt.
In addition, the innovativeness assumes Intel dwindling. Last year, Intel officially stepped down from the “tick-tock model” where every year either introduced a new architecture, or a new process. Instead, there now follows the tock a refresh with slightly higher clock speeds, but based on the same architecture and the same process. At Skylake called this refresh Kaby Lake, which even for the “tick-tock-refresh ‘model will be slightly put too late to market. Here it is not, at: reportedly comes even after Kaby Lake not a new process, but a processor based on 14nm, called Coffee Lake.
All this may be fine for laptops and mobile devices, but desktop users are obviously not happy with it, and we at Hardware.Info either. Perhaps a sound Zen-processor Intel can also shake to pay more attention to the needs of desktop users and gamers.
At Zen AMD will abandon the abnormal processor architecture that was introduced with Bulldozer and the processor goes back more like the offerings from Intel. With these changes, the smaller process and other optimizations would Zen according to AMD in performance per clock Broadwell-E should sit, which would be a giant step forward compared to the current generation of AMD processors.
So there is enough reason to be cautiously enthusiastic about Zen. Even if AMD succeeds failed to match the performance of Broadwell-E, it would even a reasonably powerful processor market quite shake. At present, Intel has in the high-end and mid-range segment, namely absolutely nothing to fear from AMD. If Zen even comes close to what AMD promises it, this will be a major positive step for consumers, because there is again talk of a degree of competition in the market for processors.
The chipsets do not yet impressed, but given the fact that very little information has been released about the highest positioned chipset, which can still be easy. In fact, chances are that we will see a traditional “AMD innovation ‘in Zen again, with the integration of more features in a high-end processor (USB and SATA ports).