Microsoft Azure introduces Graphcore’s IPUs.

 Microsoft Azure introduces Graphcore’s IPUs.

Graphcore, the AI chip startup, has announced that it’s expanding collaboration with Microsoft to implement its intelligent processing units (IPUs) on Microsoft Azure cloud. This innovation will make Microsoft the first public cloud service provider that offers IPU for machine learning workloads.

According to the partners’ reports, the testing of a sample server, equipped with eight Graphcore C2 IPU-processor PCIe cards, showed impressive results.

Based on the performance of the BERT language model used for natural language processing models pre-training during a 56-hour period, the throughput of the server running on Graphcore IPUs turned out to be three times higher compared to a conventional system. The companies also reported improvements in latency. Natural language processing models are very important to Microsoft, which is not surprising given the steadily increasing interest in cloud platforms and all kinds of voice and language services.

Graphcore is presenting its IPU technology as a GPU competitor, claiming the chip’s 100-fold performance superiority over GPUs and other AI chips in a number of specific tasks. 

Graphcore announced it is also partnering up with Dell Technologies. The software company’s DSS 8440 machine learning server, equipped with Graphcore’s IPU technology, will be demonstrated at the SC19 conference in Denver. This collaborative development uses eight Graphcore C2 IPU PCIe cards, connected via high speed Graphcore’s IPU-Link bus enabling zero latency. The partners are planning that their product would target on-premise machine learning workloads.

Meanwhile, Graphcore C2 has a quite compelling architecture. It contains 1216 IPU-Tiles, each with an independent IPU-Core executing 7296 programs in parallel. The chip has 300 MB in-Processor-Memory and 45 TB/s of total memory bandwidth. The IPU-Link chip to chip bandwidth is 320 GB/s.

The IPU runs on Poplar, a specifically designed software stack, which easily integrates with a few popular tools for developing machine intelligence apps such as the TensorFlow open source library or the Open Neural Network Exchange. By the end of 2019, Graphcore is also planning to introduce initial support for the PyTorch ML library, with full support expected in early 2020.


Are you ready to make the most of IT? Schedule a call with an expert today.