In the field of servers and data centers, which ordinary users seldom pay attention to, AMD has also set off a storm in Snapdragon with its brand-new EPYC, which is even more remarkable and unexpected.
In the middle of 20 17, AMD officially returned to the data center field with Snapdragon. The first generation product code-named Naples has powerful specifications such as 14nm technology, up to 32 cores and 64 threads, 8 channels of DDR4 2TB memory, 128 PCIe channel, independent security subsystem, SoC single-chip integrated design, and multi-generation compatible interfaces.
The little dragon reborn from the fire has attracted the attention and praise of the whole world, and even won a warm welcome from the whole server industry. Related products and solutions are becoming more and more abundant, and the market share has returned to 5% in just over a year.
Of course, all this is just the beginning. At the beginning of the birth of Zen architecture, AMD made public the road map for many years to come.
Forrest Norrod, Senior Vice President of AMD and General Manager of Data Center and Embedded Solutions Division, said in an interview with the media that AMD has formulated a three-step strategy, namely, Zen architecture in 20 17, Zen architecture in 20 19, and Zenarchitecture in 2020.
Forrest Norrod said: "Naples is a good start. It allows us to re-enter the data center market. So far, we have achieved this goal well. Rome is an important part of this plan, and we hope it will continue to lead in performance and further shorten the gap with Intel. Milan will surpass their competitors in performance. "
20 18, 165438+ 10, AMD unveiled many details of the second generation Opteron processor code-named Rome for the first time, including brand-new 7nm technology, brand-new Zen 2 architecture, up 4TB DDR4 cores, 128 threads,1.
With new technology and new architecture, the second-generation Snapdragon claims that the comprehensive performance is directly doubled and the floating-point performance is quadrupled compared with the previous generation. The CPU has never been so thoroughly improved in history.
On June 20 19, Dr. Lisa Su, CEO of AMD, was specially invited to give a keynote speech at the exhibition, during which it was officially announced that the second-generation Snapdragon server processor series was expected to be launched in the third quarter of 20 19. Now the third quarter has arrived, and the second-generation Snapdragon has arrived at the gate.
The chiplet chip design concept of the second generation Snapdragon is in line with the times. Today, when Moore's Law is gradually failing, it is unrealistic to integrate all modules on one chip, which is not allowed in terms of specifications, performance or economy. We must find another way.
AMD Zen architecture didn't blindly make a super-large core from the beginning, but split it into parts. Four cores form a CCX basic module, and two CCX basic modules form a chip, and multiple chips are packaged uniformly, so the whole processor can be customized as required. Do as much as you want, and at the same time reach a very ideal level in terms of cost and yield.
Snapdragon II went further. It can not only combine up to eight chips to realize the super-large scale of 64-core 128 threads, but also deliberately separate the I/O input/output part into another chip and adopt a more mature 14nm process, and then connect them in series through the Infinity Fabric bus with high bandwidth and low delay, so that different modules form an organic whole.
The benefits of this design are obvious. On the one hand, product specifications can be easily enlarged and strengthened, and can also be adjusted at will according to needs to meet the needs of different levels. On the other hand, it can effectively simplify the difficulty of product design, improve the yield of production, reduce various costs, and ensure the rapid and smooth launch of products.
Then, after adopting more advanced design concepts and more advanced configuration specifications, what is the performance of the second-generation Snapdragon? AMD has also made many public demonstrations, which can surprise the audience every time.
During the technical meeting of 1 1 last June, one-way Rome Snapdragon singled out two-way Intel flagship Xeon Platinum 8180m: the 7 nm process was 14nm process, and the 64-core 128 thread was 56-core 12 thread, the largest.
Specification spike, at the same time, in the industry standard performance test C-Ray project, a Roman second-generation Snapdragon took 28. 1 second to complete, while two Xeon 8 180M took 30.2 seconds, and two directly spiked!
During the Taipei Computer Show, the second generation Xiaolong once again demonstrated his magical skills. According to Dr. lisa su, the performance of the second generation Snapdragon in different applications can be as high as 2-4 times that of the first generation.
In terms of performance PK, AMD Rome II Snapdragon produced two top 64-core 128 threads, while Intel is a newly released Cascade Lake II scalable Xeon, and its process architecture has not changed at all. The top model Xeon Platinum 8280 is still 28 cores and 56 threads, but the frequency is 200MHz higher than the previous generation.
Results The second-generation Snapdragon achieved 19.88 ns// day, while the second-generation scalable Xeon achieved only 9.7 1 ns// day, and AMD was twice as advanced.
In fact, what really proves the strength of the second generation of Snapdragon is not only how dazzling the specifications and parameters are, how attractive the performance demonstration is, but also the attitude of the industry. The second generation Snapdragon has not been officially released, and even in the initial stage of customer verification, it has gained unprecedented welcome.
Especially in the field of supercomputers, which can be called the "crown" of high-performance computing, the second-generation Snapdragon has achieved a historic leap, not only achieving a breakthrough, but also stepping into the top ranks in the world.
Germany announced early that the new generation of supercomputing Hawk will be equipped with a second-generation Opteron processor with as many as 10000 6464 cores and 128 threads, with a total of 640,000 cores and1280,000 threads. The peak computing performance can reach 24.06PFlops, or 240 million floating-point operations per second.
Cray's first ExaScale new generation supercomputing platform, Shasta, immediately joined the support for Snapdragon.
The US Department of Energy has a special liking for Xiaolong. Based on Cray's new platform, two new supercomputers Perlmutter and Frontier have been built based on customized versions of Snapdragon. In particular, the latter also added AMD Radium Dragon Instinct Calculation Card. It is estimated that the floating-point performance can reach1.5eflops (1500 million times), which is not the first in the world but also the top level in the world.
At this rate, the future super-TOP500 list will undergo earth-shaking changes, and Snapdragon will surely become the backbone that cannot be ignored.
The "new arrival" of a generation of Xiaolong has attracted almost all corporate customers around the world. The domestic BAT Big Three has been introduced in the first time, and Amazon, which is difficult to be embarrassed, has been deployed on a large scale. Dell, once dismissive, was finally persuaded to adopt it completely. The prospect of the second generation of Snapdragon can be said to be infinitely bright, especially the epic leap of supercomputing.
As we said before, AMD's Snapdragon product line has made a robust roadmap from the very beginning, and implemented it step by step, with a time error of no more than one quarter, and also referred to the progress of competitor Intel.
It is interesting to note that the original rival of AMD 7nm Rome II Snapdragon is not the newly released Cascade Lake II scalable Xeon, but the new generation 10nm craft ice lake, which is expected to have better performance. However, I didn't expect the Intel 10nm process to shrink again and again. This year's first batch of products can only be used for low-power thin and light books, and desktop platforms cannot be boarded.
According to the officially disclosed roadmap, Cooper Lake, Intel's next-generation Xeon code, will still be 14nm process and old architecture, and there will be no essential changes. Snapdragon's original rival, 10nm ice lake, will have to wait until next year at least.
As for AMD, next year we will see the third-generation Snapdragon, code-named "Milan", based on 7nm+ new technology and Zen 3 new architecture, which is progressing as planned; Then it will be the fourth generation of Zen 4' s new architecture, and its code name was announced as "Genoa" early. We are making every effort to design, and the leading edge will be bigger and bigger. ...
It can be said that AMD Snapdragon ushered in the best opportunity in history at the right time and place, and such a dazzling and non-stop combination boxing continues to be played. What else can I say? AMD Snapdragon's original goal of returning to the top is believed to be only a matter of time.