We are talking about multi-core CPUs from global manufacturers.
/ photo PD
48 cores
At the end of 2018 Intel Cascade-AP architecture. These processors will support up to 48 cores, have a multi-chip layout and 12 channels of DDR4 DRAM. This approach will provide a high level of parallelism, which is useful in processing big data in the cloud. The release of products based on Cascade-AP is scheduled for 2019.
on 48-core processors and in IBM with Samsung. They create chips on the architecture . The new devices will support the OpenCAPI 4.0 protocol and the NVLink 3.0 bus. The first will provide backward compatibility with POWER9, and the second will speed up data transfer between the components of the computing system up to 20 Gb / s. The POWER10 is also known to feature new I/O technologies and improved memory controllers.
Initially, the chips were supposed to be manufactured at GlobalFoundries using a 10nm process technology, but then the choice was made in favor of TSMC and 7nm technology. The development is planned to be completed between 2020 and 2022. By 2023, the corporation will also release POWER11 chips manufactured using a 7nm process technology with a transistor density of 20 billion units.
On , 48-core Intel solutions are three times faster than AMD counterparts (with 32 cores). As for the POWER10, nothing is known about its performance yet. But that the new generation of processors will find application in the field of analytics and big data analysis.
56 cores
Such chips were recently announced by Intel - they will be manufactured using the 14-nm process technology. They support 3D Xpoint-based Optane DC memory modules and have patches for Specter and Foreshadow vulnerabilities. The new devices come with 12 memory channels and a range of built-in accelerators for cloud computing, AI and ML systems, and 5G networks.
The flagship model with 56 cores will be called the Platinum 9282. The clock speed will be 2,6 GHz, with the possibility of overclocking to 3,8 GHz. The chip has 77MB of L3 cache, forty PCIe 3.0 lanes, and 400W of power per socket. The price of processors starts from ten thousand dollars.
Developers that Optane DC will reduce the reboot time of computing systems from several minutes to several seconds. Also, the new chip will allow you to operate a large number of virtual machines in the cloud. A 56-core processor is expected to reduce the cost of maintaining a single VM by 30%. However, experts that the new processors are, in fact, an updated version of the Xeon Scalable. The microarchitecture and clock frequency of the chip remained the same.
/ photo CC BY-SA
64 core
Such a processor at the end of last year at AMD. We are talking about the new 64-core Epyc server chips based on the 7nm process technology. They should be presented this year. The number of DDR4 channels will be eight at a frequency of 2,2 GHz, and 256 MB of L3 cache will also be added. The chips will 128 lanes of PCI Express 4.0 instead of version 3.0, which will double the throughput.
But a number of Hacker News residents that productivity growth is not always "on hand" for potential users. Following the overclocking of capacities, the price of processors also increases, which can reduce consumer demand.
Developed a 64-core processor and Huawei. Their Kunpeng 920 chips are ARM server processors. It is manufactured by TSMC using a 7nm process. TaiShan servers have already been equipped with new devices with a clock frequency of 2,6 GHz, support for PCIe 4.0 and CCIX interfaces. The latter are designed to work with big data and applications in the cloud.
Huawei processors have already shown a performance boost of 20% in tests with TaiShan servers. In addition, the memory bandwidth has increased by 46% compared to previous products of the corporation.
Total
In general, we can say that the competition in the server chip market in 2019 will be high. Manufacturers are adding more cores, equipping processors with support for new data transfer protocols, and trying to make products multitasking. Due to this, data center owners have more opportunities to choose solutions suitable for a specific type of workload and for specific tasks.
Additional materials from our Telegram channel:
Source: habr.com