The POWER9 is a brand-new processor constructed with one intention in mind: being faster than Intel when working with AI. In IBM’s new AC922 server, two POWER9 chips are combined with up to six NVIDIA Tesla V100 GPUs. The compute units work together to crunch through the massive datasets required for neural network training.
The POWER9 chips have over 8 billion transistors and are engineered on a 14nm platform. Options will be available from 16 to 44 cores, depending on the compute intensity of the target workload. Each chip supports up to three GPUs and 16 DDR4 memory slots, allowing for a maximum of 1TB.
IBM’s also improved the linking technologies over POWER9’s Intel rivals. The company claims POWER9 is the first processor series to feature PCIe 4.0. This allows data to move between the CPU and GPU up to twice as fast as the current PCIe 3.0 standard, speeding up AI training when both resources are used in tandem.
READ NEXT: 30% of business apps to gain AI enhancements by 2019
The real test for IBM will be when new Intel and AMD-powered x86 chips show up with similar AI performance boosts. IBM claims that POWER9 currently offers significantly higher link speeds than its x86 competition. It hasn’t published any real-world performance comparisons though, instead focusing on what POWER9 can achieve for customers.
“We’ve built a game-changing powerhouse for AI and cognitive workloads,” said Bob Picciano, SVP of IBM Cognitive Systems. “In addition to arming the world’s most powerful supercomputers, IBM POWER9 Systems is designed to enable enterprises around the world to scale unprecedented insights, driving scientific discovery enabling transformational business outcomes across every industry.”
IBM’s already signed up several significant clients to purchase POWER9 servers. Google will use some POWER9 units in its datacentres and the U.S. Department of Energy has chosen the platform for its upcoming “Summit” and “Sierra” supercomputers. More server models will be released over the next year to target specific workloads, with options for increased scalability and higher memory capacity confirmed to be in the pipeline.
POWER9 systems will be compatible with most popular deep learning frameworks, including Caffe, Chainer and TensorFlow. IBM said there will be “demonstrable performance improvements” with any of these frameworks, compared to the industry competition. This will allow developers to create more sophisticated AI models in shorter spans of time.
