Since taking office as Intel's eighth CEO in February 2021, Pat Gelsinger has gone from CEO of hybrid cloud and computing software to CEO of the world's largest "core company." As the world's largest chip designer and manufacturer, Intel has occupied the largest market share in the global PC and server chip market since the launch of the Pentium processor in 1993, and has continued to innovate in communications, Internet of Things, networking, artificial intelligence, 5G, autonomous driving and other fields in recent years. However, Intel, which missed the mobile era like Microsoft, can ride the "cloud" in the era of cloud computing?
At the 2021 Hangzhou Yunqi Conference hosted by Alibaba Cloud, Intel comprehensively demonstrated the major innovations since Kilger became intel CEO, as well as the in-depth cooperation with ultra-large-scale cloud service providers such as Alibaba Cloud. It can be said that the eye-catching "core" results have made the industry full of expectations and confidence for Intel in the era of hybrid cloud and computing! From hybrid computing to future data centers to hardware innovations such as edge computing, as well as software innovations such as AI, vision, and hybrid cloud, we can see the significant changes that Kissinger has brought to Intel, and Intel has once again demonstrated the "wisdom" and determination to start from the "core".
<h1 class="pgc-h-arrow-right" data-track="8" > hardcore innovation for digitalization</h1>
At the 2021 Hangzhou Yunqi Conference· At the Intel Forum, Chen Baoli, Vice President of Intel Marketing Group, China Data Center Sales General Manager, comprehensively demonstrated Intel's innovative architecture and products that enable hybrid cloud and computing with full stack. The so-called "full-stack empowerment" is to comprehensively improve the performance of various scenarios of hybrid cloud and computing from the basic level of chip architecture, and even empower new scenarios.
Chip architecture determines the software and hardware architecture, Intel's chip architecture will be updated every year, but this year's strength is very large, can be called a generation of chip architecture of the big migration. As we all know, Intel founder Gordon Moore proposed Moore's Law in 1965, and in the era of hybrid cloud and computing more than 50 years later, the industry generally believes that Moore's Law has reached the ceiling, but Intel does not think so. Chen Baoli stressed that Moore's Law is not a physical law, but a goal proposed by Gordon Moore to Intel and the entire chip manufacturing industry at the beginning, that is, to believe that advanced manufacturing technology can double the number of transistors on integrated circuits every 18 months, and the corresponding performance of processors will double every two years. Now, Intel's architects are constantly thinking and exploring, looking at chip design through different dimensions and improving chip efficiency.
Today's new applications for the hybrid cloud and computing era, with computing power tripling in 18 months, can't keep up with the pace of new applications. Therefore, Intel thinks and explores from five dimensions, such as process, connection, memory, software, and architecture, and promotes each dimension to quadruple the computing power in the next three to four years, then the five dimensions are the 5th power of 4, that is, 1000 times the performance improvement. This is Intel setting itself a new goal, that is, through multi-dimensional advancement, to create more computing power, to support a variety of new applications in the future.
How to achieve a thousand times more computing power in the next three to four years? This is the hybrid computing architecture. Intel from the traditional CPU's Scalar tensor computing engine, to today's AI needed Matrix computing engine and more types of computing required Vector vector computing engine and Spatial spatial architecture computing engine, etc., the formation of heterogeneous computing engine architecture, coupled with different processes, packaging technology, and for different computing workloads with the appropriate cache, memory and connection, you can form many different types of computing products under the hybrid architecture. Meet every need. When a single computing core is spliced together to form a multi-core CPU, each computing core can assume different functions, which is a new hybrid computing design concept.
Under the new hybrid computing architecture design concept, Intel has made great improvements and performance improvements to the well-known X86 chip, intel also known as Performance Core, especially the new New Matrix Engine for general AI computing, which can be used for common artificial intelligence inference calculations, such as finding the picture you want to find in tens of thousands of maps can be implemented in seconds with the new Performance Core. Eficient Core is an efficient, high-performance core optimized for multi-threaded and multi-core computing, which can provide greater efficiency and performance for lightweight parallel multitasking computing in cloud computing environments, such as multi-document uploads in cloud backups. Alder Lake, a product that combines Performance Core and Advanced Core, is geared towards the next generation of PCs and will be available in the second half of this year.
In terms of next-generation data centers, IPUs (Infrastructure Processors) are Intel's latest processors to help cloud service providers reduce costs and improve CPU performance. To put it simply, IPU is to transfer the infrastructure functions of cloud service providers to the IPU, thereby releasing the maximum computing power of the CPU. IPU can also promote the pooling of basic resources, so that CPU, GPU, memory, etc. are pooled under the scheduling of IPU, thereby maximizing the service capabilities and cost reduction of cloud service providers, Mount Evants is Intel's first ASIC IPU product, but also the result of cooperation with a top cloud service provider.
Ponte Vecchio is Intel's new GPU for high-performance and AI workloads, which contains 1,000 trillion transistors and existing engineering chips that can reach 45 trillion floating-point computations per second, which is already the industry's ceiling. Intel has also launched the next generation of Intel Xeon Scalable processors for data centers - Sapphire Rapids, which for the first time uses multi-stitching connections, brings more cores, and adds many acceleration engines, including artificial intelligence acceleration engine instruction set AMX and data streaming instruction set DSA, etc. DSA can help cloud service providers manage all CPU cores more efficiently and improve utilization. Sapphire Rapids also integrates a number of innovative technologies that bring more innovation, more computing power, and more breakthroughs to the data center.
<h1 class="pgc-h-arrow-right" data-track="41" > hardcore cooperation with hyperscale clouds</h1>
While rethinking Moore's Law, Intel is also returning to the essence of chips — using hardware engineering to solve computing power and inclusive computing problems. Therefore, working with the end customer to think about the most pervasive computing problems is an important way to innovate. Intel cooperates with ultra-large-scale public clouds such as Alibaba Cloud to return to the essence of innovation and re-engineer the "core" direction by studying ultra-large-scale computing problems.
Intel Optane Persistent Memory is a new species that is born to solve the computing speed and persistent storage of the data center. Intel Optane Persistent Memory is characterized by its large capacity, economy, and durability, which can bring new improvements to a variety of data center computing scenarios, including analytics, databases and in-memory databases, artificial intelligence, containers, and more. Guoan Wu, Manager of the Engineering Department of Intel China Optane Innovation Center of Excellence, and Xianghua Tang, an expert in Alibaba Cloud elastic computing products, introduced the joint efforts of the two parties to promote cloud infrastructure and cloud application innovation with Intel Optane persistent memory at the Intel Forum at the 2021 Hangzhou Yunqi Conference.
Intel Optane persistent memory can serve as an on-cloud infrastructure for application innovations for all customers on the cloud, such as leveraging persistent memory to expand memory capacity in virtual machines or containers or accelerate storage. Wu Guoan introduced Optane persistent memory for cloud innovation from two directions. The first is that for scenarios where computing or extra overhead is particularly large, such as network overhead, etc., you can use Optane persistent memory instead of memory, so that there is no obvious loss in performance, but TCO is much lower. The second direction is the scene where the IO account for a very large proportion, that is, the storage proportion is large, because the basic performance of Optane persistent memory is dozens of times the performance of SSDs, so it can be accelerated by Io with Optane persistent memory, thereby greatly improving the overall performance.
For the first direction, using Optane persistent memory to accelerate the Redis database is a typical scenario. Redis was named the Top 1 Non-SQL in-memory database in 2021, and the database has a large number of applications among many customers. Replacing memory with Optane persistent memory can significantly reduce TCO with little loss of Redis performance. Of course, it is also possible to completely replace DRAM memory without Optane persistent memory, but to perform certain ratios according to different business scenarios, and to perform hierarchical processing of data, such as layering MemKeyDB. In addition, Alibaba's self-developed distributed caching system Tair also uses Optane persistent memory, which effectively reduces costs and expands memory.
For the second direction, accelerating KV storage with Optane persistent memory is a typical scenario. KV storage is the basis of many applications on the cloud, for example, the parameters of Alibaba DeepRec's recommendation system use KV storage based on Optane persistent memory to achieve storage acceleration. There are also open source PMEMKV storage supported by Intel, KVDK development tools, and RocksDB, which is open sourced by Facebook.
Tang Xianghua introduced Alibaba Cloud's cloud innovation practice based on persistent memory from the product landing level. Alibaba Cloud has long been involved in exploring persistent memory, Alibaba Cloud is the world's first manufacturer to launch a cloud server based on persistent memory, as early as last October released the first persistent memory instance. After last year's precipitation and polishing and iteration of the underlying hardware architecture, Alibaba Cloud has carried out a comprehensive upgrade iteration of persistent memory instances this year - the latest persistent memory instances are based on Alibaba Cloud's third-generation DPCA architecture and Intel's latest second-generation persistent memory BPS, which has doubled the IO capabilities and increased computing power by more than 40% compared with the previous generation of products.
At the product level, Alibaba Cloud has made bold innovations this year, launching a high-end local disk instance product i4p based on persistent memory, which has achieved fundamental performance improvements compared with traditional NVME local disk instances, especially the read and write latency of high-end local disks can be as low as 170 nanoseconds, while the traditional NVME is about 10-30 microseconds, so the performance is increased by nearly 100 times. In addition, the single disk IOPS of high-end local disks is as high as more than 1.4 million, and the throughput of single disks is as high as 9GB/s, which is also several times higher than NVMe. Due to the absolute leading performance of high-end local disks, it is a great blessing for customer application scenarios with heavy IO - the performance bottleneck problems caused by the lack of performance of NVMe SSDs have been effectively solved on high-end local disks, such as 2.5 times the performance of RocksDB database, 2 times the performance of Click House database, 2-3 times higher recovery time of NSQ message middleware services, and effective suppression of inertia and jitter of heavy IO applications.
In addition, Alibaba Cloud has continued to explore the memory characteristics of persistent memory, and launched memory-based memory instances r7p and re7p in parallel with i4p. For customers who need to consume a lot of memory, replacing memory with persistent memory can immediately reap the cost benefits. Because large-capacity content is very expensive, and no other cloud service provider in the industry can provide instances with small specifications and large-capacity ratios, it is impossible to expand memory indefinitely, but can only compromise between performance and cost. Alibaba Cloud sees that persistent memory can solve this problem, especially some applications have undergone specific transformations, which can allow customers to obtain pure cost benefits under the premise of almost lossless performance, such as redis, Parameter Server and other applications can reach more than 50% price-performance improvement and application deployment is very simple, there are already many Alibaba Cloud customer applications running on persistent memory.
In the face of endless applications in the cloud, Intel persistent memory is also constantly innovating: the third generation of products is in the process of development, which will have a great improvement in performance; the fourth generation of products uses a CXL interface, which can be directly plugged into the PCI-E bus and expanded with memory. Large memory requirements will be fully freed from the growing number of processor cores, while increased network performance will require higher performance, higher-life storage—these are innovative scenarios where persistent memory plays an important role. In order to promote persistent memory, Intel has also launched a corresponding application development programming book to help more developers innovate with persistent memory.
<h1 class="pgc-h-arrow-right" data-track="42" > wisdom in this core, soft and hard</h1>
The DeepRec Massive Sparse Model Training Engine is Intel's leading AI engineering system in partnership with Alibaba, a tensorFlow-based recommendation engine. As we all know, there are a large number of recommendation systems in Alibaba's business scenarios, and the feature values of the recommendation systems are very sparse, especially suitable for Intel technology. As a unified training engine for Alibaba Group's sparse scenarios, DeepRec is jointly built by multiple teams of the Group, and has supported Taobao's core businesses such as search, recommendation, and advertising since 2016, and also provided services for Alibaba Cloud customers' recommendation business.
Since the characteristics of Alibaba's recommendation system are often very large, reaching the super-scale scale of tens of billions, hundreds of billions, trillions or even 10 trillions, how to achieve efficient business landing based on Alibaba's machine learning platform PAI and Intel technology has become a challenge that both sides must jointly solve. Wei Bin, Director of Ultra Cloud Computing Software Architecture at Intel China, and Huang Boyuan, Senior Product Expert of Alibaba Cloud Intelligence, introduced the efforts made by both parties around the DeepRec engine at the Intel Forum of the 2021 Hangzhou Yunqi Conference.
Huang Boyuan introduced that the recommendation system has great commercial value, but the deep learning model of the recommendation system faces several major challenges: first of all, the processing of the model eigenvalues, because the eigenvalues of the recommended model change very quickly, it needs dynamic processing and cannot wait a week or a month like traditional deep learning training to train a model; in addition, due to the ultra-large-scale model, in the training process, it will encounter a series of problems such as building a distributed framework, runtime and compilation optimization, graph optimization, operator optimization, etc. All of this optimization requires deep cooperation between the upper layer software and the underlying hardware and the operator library provided by the hardware vendor, so the PAI platform chose to work with Intel to optimize DeepRec.
Wei Bin introduced that Intel and Alibaba PAI machine learning platform team cooperated to complete a large number of optimizations for DeepRec, including a large number of operator optimizations, graph optimization, runtime optimization, compilation optimization, and deep customization optimization based on hardware, supporting high-performance distributed training, and excellent performance in the training of sparse models. For example, in terms of operator optimization, the typical operator optimization for the sparse model makes full use of AVX512 and BF16 instruction acceleration, and also blocks the large matrix to improve cache hit rate, reduce the memory overhead, and achieve operator acceleration by 2-6 times; in terms of framework and library optimization, the latest Intel oneDNN acceleration library is integrated to achieve thread pool scheduling optimization, obtain the best performance, and optimize based on PEME persistent memory.
At present, Intel is working with Alibaba to promote the open source of DeepRec, thereby benefiting the entire ecosystem, so that more customers can get the same performance as Alibaba's recommendation system, and thus take the lead in the development of digital business.
【Full Text Summary】From the 2021 Hangzhou Yunqi Conference Intel Forum, it can be seen that Intel is determined and capable of carrying out low-level hardware and chip-level innovation for future digital scenarios. Through cooperation with the world's top cloud service providers such as Alibaba Cloud, Intel has opened up the ideas and directions of innovation from the "core", and at the same time, in the actual ultra-large-scale business polishing, it has tempered the integrated innovation and optimization ability from the underlying chip to the upper-layer application software, and found a way to win the hybrid cloud and computing era. So, in the era of hybrid cloud and computing, Intel can expect it! (Text/Ningchuan)