Graphcore, the U.K. AI chip developer, is expanding collaboration with Microsoft to offer its intelligent processing units on the Azure cloud, making Microsoft the first large public cloud vendor to offer the IPU designed for machine learning workloads.
Azure support for IPUs is the culmination of more than two years of collaboration between the software giant and the AI chip startup. Early development work has focused on enhancing natural language processing (NLP) and machine vision models running on the IPU.
The partners reported model performance and accuracy milestones based on the BERT language model used for NLP pre-training. Models were trained in 56 hours using one IPU-based server equipped with eight Graphcore C2 IPU-processor PCIe cards. The result was a three-fold throughput increase for BERT inference along with latency improvements, the partners said.
“Natural language processing models are hugely important to Microsoft—to run our internal AI workloads and for our AI customers on Microsoft Azure,” said Girish Bablani, corporate vice president for Azure Compute.
Graphcore has been positioning its IPU as a GPU competitor as well as an AI chip customized for machine intelligence training and inference applications. The startup has claimed up to a 100-fold performance boost over GPUs and other AI chips.
The chipmaker differentiates its parallel processing design by handling emerging machine intelligence workloads, including spatial AI and simultaneous localization and mapping models used in robotics applications.
Further, the AI chip retains full machine learning models with the processor, meaning a 16-processor architecture with a custom server link could handle more than 100,000 independent programs, “all working in parallel on the machine intelligence knowledge model,” the startup claimed in a Nov. 14 blog post announcing the Microsoft Azure partnership.
Along with early access customers focused on automation and financial applications, Graphcore also said it is collaborating with Dell Technologies to demonstrate its IPU technology on the hardware vendor’s DSS 8440 machine learning server.
Graphcore said the combination could deliver more than 1.6 petaflops of machine intelligence processing using eight Graphcore C2 IPU PCIe cards, each with two IPU processors and connected via high speed Graphcore’s IPU-Link technology running on a standard 4U chassis. The combination would target on-premise machine learning workloads, the partners said.
Together they will be demonstrating the technology at SC19 at Dell Technologies’ booth (#913),
The IPU runs on a software stack labeled Poplar. The result, Graphcore said, is a “graph tool chain specifically designed for machine intelligence” applications. Poplar works with the TensorFlow open source library and the Open Neural Network Exchange to help developers using existing machine learning tools and models.
The startup said initial IPU support for the PyTorch machine learning library will be available by the end of this year, with full PyTorch support coming in early 2020.