The new benchmark of high-end processors Power10 refers to the ‘memory wall’
Power10 CPU is the highest-end general-purpose processor in the current industry. It aims to provide users with a platform to meet future computing needs and set a new benchmark for enterprise-level computing. The Power10 processor is a commercial processor built using 7nm process technology. Compared with the POWER9 processor, it is expected to increase 3 times in terms of energy efficiency, workload capacity and container density.
Power10 processor is the latest generation of central processing unit products in the POWER architecture series of server products. Several innovative technologies have attracted attention. Open memory interfaces, ultra-high memory bandwidth, and lower minimum latency. Power10 further breaks through the ‘memory wall’. Restrictions set a new benchmark for high-end processors.
The Power10 processor uses an advanced 7nm process in its manufacturing process. It is the first commercial chip to use the 7nm process. It has obtained and applied for as many as hundreds of patents. The Power10 processor can integrate up to 15 cores, all cores support SMT8, that is, each core supports eight threads in parallel, and a single chip can provide up to 120 threads. In addition, the L2 cache of each core is 2MB, and the L3 cache of each processor is 120MB, and the low-latency NUCA cache management technology is adopted.
In terms of packaging technology, due to the advanced 7nm process, the core area is smaller than the POWER9 processor, only 602mm², but it integrates up to 18 billion transistors, and each slot supports both single-core and dual-core In the form of packaging, the 18-layer metal stack traces far exceed the current industry standards. The low-level metal traces are responsible for the internal circuit interconnection, and the high-level metal traces are responsible for the power supply distribution, clock and off-chip signal interconnection, which can better ensure the RAS of the chip. characteristic.
Based on the focus on performance and efficiency design, Power10 is expected to increase the energy efficiency of each socket processor by 3 times. Compared with POWER9, it can increase the workload capacity under the same power consumption. Compared with server products equipped with POWER9 processors, this expected increase in capacity can increase the workload and container density of Power10-based systems by 3 times, which will affect multiple data center attributes and improve efficiency. Reduce costs (such as space costs, energy costs, etc.).
In terms of instruction set, the Power10 processor through advanced enhanced processor core architecture and built-in embedded matrix math accelerator (Matrix Math Accelerator), making each socket based on FP32, BFloat16 and INT8 calculation of AI inference speed compared to POWER9 processing The devices are 10 times, 15 times and 20 times faster respectively, thereby improving the performance of AI inference workloads and integrating them into commercial applications to achieve faster insights.
The breakthrough PowerAXON technology breaks the memory boundary of a single server and brings hardware-integrated memory call opportunities for users' memory-intensive workloads; and simplifies the management of the SaaS layer, allowing users to focus more on their application loads And business innovation.
The Power10 processor enhances this function by providing pooled or clustered physical memory on Power10-based systems through various configurations. With the breakthrough new technology of Memory Inception, the Power10 processor allows any Power10 processor-based servers in the cluster to access and share each other's memory, thereby creating a PB-level memory cluster and providing a cost-effective platform for massive memory-sensitive applications.
In terms of I/O, Power10 integrates a set of memory controllers called Open Memory Interface, which can support GDDR, DDR and SCM memory modules, which is more flexible and can cope with a wider range of memory. Application scenarios. Among them, the support for DDR has achieved the latest DDR5 memory, this set of interfaces can provide up to 1TB/s of ultra-high bandwidth, and the theoretical peak bandwidth of each slot can reach up to 410 GB/s.
In addition, because Power10 uses the new core of the new transistor, it can enable the system memory to achieve a higher transmission band, and supports the PCIe 5.0 bus. The PCIe bandwidth is 2 to 4 times higher than that of PCIe 4.0. Power10 processor integrates PowerAXON interface, unique high-speed signal interconnection technology can provide up to 1TB/s signal connection, of which 16 slots are interconnected through SMP, memory and I/O are connected to OpenCAPI interface, and adjacent The server system memory forms a memory cluster.
For a long time, POWER is an advanced computing platform that supports flexible deployment of cloud and local workloads. In the Chinese market, based on the advantages of Power10, Inspur Business Machines will focus on industry clouds, and promote the implementation of public cloud in China through joint industry partners, and provide users with a ‘three high and one strong’ experience in the cloud environment, while reducing The user's overall application cost.
In the Chinese market, Inspur K1 Power server, as a high-end and high-value series of independently developed Inspur server family, fully releases high-value computing power. It is oriented to key core applications and has high reliability, high performance, high security and strong real-time data consistency. The computing power guarantee of the ‘three highs and one strong’ feature is widely used in banking, securities, telecommunications, medical, government and other industries to help users easily cope with the challenges brought by innovative business growth and application loads. It is the high reliability and reliability of key applications. High-performance security platform.
