IBM unveils 7nm Power10 processors for big data and AI analysis

0
279

IBM unveils 7nm Power10 processors for big data and AI analysis

IBM unveiled the IBM Power10 processor at this year’s Hot Chips 2020 virtual conference. As the name suggests, Power10 represents the next step in the evolution of the IBM Power architecture. IBM claims the Power10 delivers three times the efficiency of its predecessor, the Power9, which was announced in 2016.

IBM
IBM

Power10 chips have been in development for over five years – they were the first commercial 7nm processors from IBM. There are several configurations in the family, and although details have yet to be revealed, according to IBM engineer and Power10 architect William Starke, the maximum single-chip solution will include no more than 15 SMT8 cores (capable of executing up to eight threads simultaneously), and a two-chip module – will offer respectively 30 SMT8 cores. The new chips use a lot of advanced technologies: the developers have patented hundreds of solutions related to Power10.

In addition to saving energy, Power10 offers hardware-based memory encryption with approximately 40% faster performance thanks to new AES cores and enhancements such as support for homomorphic encryption methods. Power10 can also provide hardware-based container protection and advanced isolation capabilities. Through hardware innovations, Power10 will enable you to develop more attack-resistant applications with little performance degradation. In addition, the chip has received Memory Inception technology, which allows any Power10-based system in a cluster to share petabytes of memory with other systems. From an AI perspective, IBM estimates that a Power10 processor with FP32, BFloat16, and INT8 accuracy will provide 10x, 15x, and 20x the inference performance for machine learning tasks, respectively, compared to Power9. The Matrix Math Accelerator is responsible for this.

Also Read:   IBM's cloud revenue increased by almost a third

IBM Cognitive Systems Leader Stephen Leonard argues that the Power10’s reduced power consumption will improve data center efficiency and lower costs while allowing hybrid cloud platforms to do more work in a smaller area. He argues that Memory Inception will drive further savings as cloud providers can offer more capabilities (such as accelerating AI workloads) by using fewer physical servers with shared memory pools and cloud users leasing fewer resources to meet their IT needs. needs.

The Power10 will be produced by Samsung, which was announced back in 2018 – this is not surprising, given that the Korean company is a member of the IBM Research Alliance. Chip samples have already been received in several variants, and their market launch is expected in the second half of 2021.

The Power10, like its predecessors in the Power family, is open for licensing and modification by over 250 OpenPower Foundation contributors including Google, NVIDIA, Mellanox, and Tyan. The previous family of IBM Power9 processors was announced at the Hot Chips conference in 2016. Unlike the Power10, it was manufactured in accordance with the 14nm FinFET regulations and was available in 12 and 24 core versions.

One of the fastest supercomputers in the world, the Department of Energy’s Oak Ridge National Laboratory Summit, is based on Power9 and works in conjunction with NVIDIA Tesla accelerators.