EETOP public number compiled from science and technology new newspaper (Taiwan) Author: idiot water balloon
Being surrounded by enemies is not necessarily a bad thing, especially if you are stronger than any opponent. Moltke, who led Prussia to win the Germanic unification war (1864 Danish War, 1866 Austro-Prussian War, 1870-1871 Franco-Prussian War), won every battle, never lost, was not only a very popular subject, but also enjoyed a long life, rich and noble, and was recognized as the most successful professional soldier in Western history.
He famously said: "We are located in the middle of the great Powers of Europe, and our eastern and western neighbors have only one side to suffer from the enemy, and we are in the belly and back of the enemy." Moltke's greatest concern was that "the Slavic East made an alliance with the Roman West and killed the Germans in the center", which became the core concept of the german war plan in the future.
On August 4, 1914, the day britain declared war on Germany and the First World War was fully detonated, if there were onlookers, circling the war table and looking at each gambler's hand, they would think that Germany's chances of winning the game were nine times out of ten. The war situation seemed unfavorable to Germany, but Germany could concentrate most of its forces against an enemy.
This is also the strategic situation of AMD's $5.4 billion acquisition of ATI on July 24, 2006, and the strategic situation is optimistic about the outside world, holding both x86 CPU and GPU, by fusing the two, it can concentrate on attacking Intel without decent GPUs or Nvidia without x86 CPUs.
Later development originated from fusion strategic missteps, so that AMD from enjoying the middle position of the geographical advantage to both ends, whether CPU or GPU, AMD was surrounded by Intel and Nvidia all the way to 2018, like Germany in europe enjoys a "central position" convenient AMD, but self-defeating and tragically killed on both sides.
However, AMD launched a major offensive in the late autumn of this year, finally showing the results accumulated over the years, and symbolizing that AMD has the research and development energy to develop a variety of processor microarchitectures for customer needs – although it is a full 15 years late.
From near to far, one by one, carefully look at what dishes AMD has come out with this wave of offensives.
Apple's M1 Max performs a unique aesthetic of violence, and AMD also has its own set of gameplay.
Ever since Intel released the Pentium III Xeon processor with 2MB of built-in L2 cache on March 17, 1999, the industry has been full of cold jokes such as "Intel's most important business is not cpu cores, but production caches". As large-scale cache memory becomes the everyday landscape of processor die-to-die graphs, this good thing is no longer just Intel's patent.
AMD (or TSMC) actively cultivated multi-chip packaging technology is not news, at the end of May this year, it was publicly displayed through the X3D package of Ryzen 5000 processor, two 8-core CCDs (Core ComplexDie) each stacked a 64MB of tertiary cache memory (size 6×6mm, area 36mm²), so that the total capacity of the third-level cache memory from 64MB to 192MB, and brought up to 2TB/s of theoretical transmission bandwidth, AMD naturally prioritizes the introduction of high-margin server product lines, and there is a faint smell of "responding to the needs of specific mega-customers".
The EPYC "Milan-X" after eight pieces of 64MB L3 cache memory not only reached a staggering 768MB capacity of L3 cache memory (plus 804MB for L1 and L2), but also reduced memory access latency by 42 to 51%, which positively countered Intel's accusation that EPYC "caused long latency due to multi-chip architecture".
Basically, Intel's third-generation Xeon SP "Ice Lake-SP" is destined to be crushed by the total number of AMD third-generation EPYC cores and cache capacity, according to known information, this situation will almost certainly continue to the next generation of competition (SapphireRapids vs. Zen 4, which will be mentioned later), and perhaps the profit margins of Intel's data center business group will take a while to return to more than the level of personal computers, and it may even be like a changed lover, never to come back.
The first two-chip package GPU in the history of the computer industry: the Instinct MI200 accelerator
But this is not tight, for AMD, the "dual-track GPU" development route revealed by the 2020 Financial Analyst Conference, the CDNA system optimized for computing needs, has finally blossomed, allowing AMD to have hardware specifications on the book that overwhelm Nvidia's high-end GPUs.
The second-generation CDNA Instinct MI200 consists of two "Graphics Compute Die" (GCD) and eight 16GB HBM2e memory, AMD made OAM (OCPAccelerator Module) strongly implies the "customer" of the lock, I heard that it seems to be a Facebook and a book renamed What Metacosm (I really don't like this translation).
The combination of MI200 and the third-generation EPYC processor has also become the main weapon of AMD's counter-attack on high-performance computing and deep learning, and Infinity Fabrics, which "unifies all OF AMD's IP function block control methods", will also play a real value.
It is worth mentioning that AMD has recently won a lot of success in the supercomputer indicator Top500, accounting for as many as 73 units in the November list, but it is extremely similar to the e-sports pen "AMD CPU with Nvidia GPU" trend, the TOP500 AMD supercomputer, EPYC is always next to the Nvidia A100. How AMD convinces customers to replace the Nvidia solution with MI200 will also be a good subject worth observing for a long time, and it is definitely a big challenge to break through the "CUDA ecosystem block network" at the special software level.
The following table also screened Intel Xe-HPC, let everyone see how the flagship GPU acceleration cards of these three manufacturers are different.
The highlight of the end of 2021: the same dual-track Zen 4
As far as the author's impression is concerned, AMD last had a way to simultaneously develop two x86 processor microarchitectures, deducting the merger of NexGen in 1995, which was the next year after the acquisition of ATi, announcing the big-core "Bulldozer" and the small-core "Bobcat" (Bobcat), and most of the time only existed in the briefing, both of which could not be published on time, the process was difficult, directly leading to the collapse of the AMD server market front.
AMD's plan to use TSMC's 5nm process to build Zen4 is not news, but the emergence of "two cores" at the same time is extremely shocking, in addition to the Zen 4 EPYC "Geona" with up to 96 cores, there are 128 "cloud computing optimization" Zen 4c (cloud optimized) core Bergamo, saying that there is no "guarantee that there are large customers to place orders" behind this is absolutely deceptive. It also fully confirms the author's view many years ago: AMD has made it clear that it is fully locked in the data center business and no longer pursues everything.
What's more frightening is that these two will appear at the same time, both support avX-512 instruction set (unknown range) and BF16 floating-point format, and are also Intel's fourth-generation 56 core 4 chip package Xeon-SP "Sapphire Rapids" is about to face opponents, when Intel will lose most of the instruction set umbrella (AMX will be the only remaining obvious strength), this "nuclear war" is bound to be more exciting, It also makes people wonder what kind of briefing content Intel will bring out next year to count the weaknesses of AMD products.
AMD has not disclosed the specification differences of Zen 4c, it may be Zen 4 reduced cache memory capacity version, or simply a new microarchitecture (such as pre-prepared Zen 5 generation small core), but the feat of "not being able to shoot in one shot, simply firing two shots" is a major step in the history of AMD x86 processor development.
Looking back, AMD is still far from the heyday of K8
In the past 15 years, many things have changed, Intel has lost its process technology and production capacity hegemony, Nvidia has become a powerful force that is difficult to shake in high-performance computing and artificial intelligence, and AMD has been steady and steady, climbing out of the mud step by step, until the market value climbs to 70% of Intel, which seems to be "returning to pesticides" (Editor's note: returning to the glory of the internet folk language), but for readers who have experienced the heyday of AMD, they should still feel that the current AMD seems to be a long way from the glory of the past.
On the data center and personal computer market, the integrity of Intel's ecosystem, AMD is still difficult to match, in terms of high-performance computing and artificial intelligence, Nvidia makes AMD completely invisible to the taillights. Intel cut the retrain gpu product line from AMD poaching Raja Koduri, and Nvidia built a complete CPU and SmartNIC technology energy through the acquisition of Arm and Mellanox. In other words, the loss of AMD," which is "in a central location", has no advantages that are not as great as it seems, and as in the past, the cost of never making mistakes, these pressures and toils are fully reflected in The increasing gray hair of Lisa Su.
But all in all, in the past three years, AMD has managed to get out of its worst situation and begin to have the ability to face Intel and Nvidia head-on. Don't forget, AMD also has the world's largest FPGA manufacturer Xilinx, and I am looking forward to how AMD will make good use of this trump card, and continue to brew the ultimate APU "EHP". There is good reason to believe that the market competition in the next few years will only become more and more joyful, so that the author who does not learn and has no skills will never lack the material to play with the problem.
Xilinx Online Conference