At the Hot Chips 2020 conference, which was held virtually this year, IBM announced the IBM Power10. It’s the successor to the Power9 and represents the next generation of the company’s processor family. IBM claims that the Power10 delivers up to three times greater efficiency than its predecessor while at the same time delivering higher workload capacity and container density.
The Power10 was designed over five years and has the distinction of being IBM’s first commercial 7-nanometer processor. (In 2015, IBM, Samsung, and other members of IBM’s Research Alliance produced the first test chips as part of a $3 billion R&D investment.) There will be multiple configurations, and while the specifics aren’t yet being disclosed, the maximum single-chip-module offering won’t exceed 15 SMT8 cores and the dual-chip-module offering won’t exceed 30 SMT8 cores, according to IBM distinguished engineer and Power10 architect William Starke.
Besides power savings, the Power10 offers hardware memory encryption with an estimated 40% faster cryptography thanks to new AES cores and enhancements such as support for homomorphic encryption techniques. The Power10 is also designed to deliver hardware-enforced container protection and isolation capabilities co-optimized with its firmware, enabling the prevention of other containers in the same virtual machine from being affected by a single intrusion.
The Power10 processor features dynamic execution register control, meaning users can design apps more resistant to attacks with negligible performance loss. Beyond this, it ships with what IBM calls “memory inception,” which lets any Power10-based system in a cluster to share memory with other systems. Meanwhile, on the AI side of the equation, IBM says it expects the Power10 processor to achieve 10 to 20 times better performance for enterprise AI inference tasks compared with the Power9.
IBM cognitive systems GM Stephen Leonard asserts the Power10’s reduced power consumption will drive datacenter efficiency and reduce costs while allowing hybrid cloud setups to achieve more work in a smaller footprint. Memory inception will driver further savings, he claims, as cloud providers offer more capabilities (such as AI workload acceleration) using fewer servers with pools of shared memory and cloud users lease fewer resources to meet their IT needs.
As announced back in 2018, Samsung will manufacture the Power10. Starke says it’s already being sampled in multiple system offerings and will likely become available in the second half of 2021.
Like previous Power processors, the Power10 is open for licensing and modification by the OpenPower Foundation’s over 250 members, including Google, Nvidia, Mellanox, and Tyan.
IBM’s last Power processor family, the aforementioned Power9, was announced at the 2016 Hot Chips conference. As opposed to the Power10, it’s manufactured on a 14-nanometer FinFET process, and it comes in 12- and 24-core versions. Summit at the U.S. Department of Energy’s Oak Ridge National Laboratory, one of the fastest supercomputers in the world, is based on Power9 combined with Nvidia Tesla GPUs acting as accelerators.