Founded in 1995, Marvell is headquartered in Santa Clara, Calif., with R&D centers in the United States, Israel, India, Germany, and China. It employs more than 5,000 people, more than 10,000 patents worldwide, and $2.9 billion in fiscal year 2019. It is one of the most respected semiconductor companies. Its core business covers three major areas of storage, network and computing. In particular, it has more than 15 years of high-performance multi-core CPU R&D experience. It has shipped billions of units and is in ARM.serverThe processor has also been worked on for generations.
In 2017, Marvell acquired the US fabless semiconductor company Cavium for about $6 billion to acquire the latter's ARM/MIPS product line, including the ThunderX2.
ARM architecture processors have unique advantages such as energy efficiency, low cost, open architecture, and extensive ecology. Since 1990, the cumulative shipments have exceeded 130 billion, and they have been accelerating. In 1991-2013, they spent 22 years shipping 50 billion. In 2013-2017, 50 billion units were shipped in only 4 years, and 100 billion units are expected to be shipped in 2017-2021.
It is manufactured by TSMC 16nm process, optional BGA integrated package, LGA independent package, based on 64-bit ARMv8.1 fully customized out-of-order execution core, up to 32 physical cores, 128 logical cores (4 threads per core), frequency regular Mode 2.5GHz, acceleration mode up to 3.0GHz, 32KB data and instruction cache per core, 256KB L2 cache, shared 32MB L3 cache.
Technical supportARM virtualization extension, RAS features, ARM TrustZone security protectionThe expansion provides up to 56 PCIe 3.0 (14 controllers), two SATA 6Gbps, and two USB 3.0.
Supports two-way parallelism,byCCPI2 coherent businterconnectionBandwidth up to 600Gbps for cross-processor cache coherency.
Memory,Supports up to eight DDR4 per channelThe frequency is up to 2666MHz and the maximum capacity of the two channels is 4TB.
ThunderX2 will be available in more than 40 different configurations, but no specific power consumption has been announced.
Marvell emphasizes that ThunderX2 outperforms Intel Xeon in terms of core count, memory bandwidth, IO channel, and price/performance ratio, while far exceeding AMD Opteron in terms of memory bandwidth and application throughput. In many cloud benchmarks, performance is comparable to 20 core 40. Threaded Intel Xeon Gold 6148 can be 10-100% higher.
The official also gave some test items of some authoritative media, Java, integer performance, memory bandwidth are very eye-catching.
According to Marvell, ThunderX2 can be used in high-performance computing, cloud computing, edge computing and many other areas to meet the highest demands of computing, storage, and networking workloads, and to accelerate the mainstream of ARM servers by providing higher levels of computing performance and ecosystem. Adoption and deployment within cloud, high-performance computing data centers.
At the same time, ThunderX2 has already had many application examples. For example, the supercomputer "Stra" of the US Department of Energy's Sandia National Laboratory is equipped with 145,152 ThunderX2 cores. It is currently the most powerful ARM supercomputer, mainly used for US nuclear weapons research.
The Los Alamos National Laboratory in the United States uses ThunderX2 cores to improve the availability of supercomputers, rather than pursuing peak performance.
The French Agency's CEA supercomputing system is also based on ThunderX2 for nuclear energy and defense related.
Ecologically, the Marvell ARM server platform has also been supported by many OEMs and platforms, IHV hardware vendors, operating systems and firmware, middleware, applications and tools.
Gopal Hegde, vice president of Marvell and general manager of the server processor department, said in an interview during the Taipei Computer Show.ARM server has a unique advantage in the cloud, in the existing huge AndroidMobile phoneOn the basis of ecology, it is possible to smoothly realize the calculation and rendering of the ARM cloud by means of the consistency of the platform and architecture. At the same time, with the improvement of the software and hardware ecosystem, the enterprise data is getting smoother from the x86 server to the ARM server, and Marvell will also be smoother. Provide wholehearted help.
At the same time, Marvell plans to launch a new generation of ARM server platforms every two years, each time achieving a performance improvement of 2 times or more.
He also revealed thatARM predicted two years ago that in the next five years, 20% of the world's servers will be based on the ARM architecture, and the proportion in the Chinese market may be higher because China dares to try new things.