In November, we announced our partnership with Graphcore to provide their Intelligence Processing Units (IPUs) on our cloud platform. Additionally, we are also selling the Dell EMC DSS 8440 machine intelligence servers with Graphcore IPUs. Graphcore IPUs present a unique opportunity for those looking to meet the computational demands of machine learning.
Recently, at the NeurIPs 2019 conference, Graphcore released additional benchmark results showing the overall performance of their IPU processors against some of today's GPUs. These benchmark results clearly show that IPUs are completely different from CPUs or GPUs and have been designed from the ground up to deliver state of the art performance on current machine intelligence models for both training and inference.
Even beyond BERT and ResNext benchmarks, Graphcore IPUs are accelerating training and inference models for other workflows. In the video below, Alex Tsyplikhin details the impressive speedups that Graphcore's finance customers have reported using IPUs for LSTM inference and MCMC model training.
Graphcore has released public examples of their benchmarks so that customers can replicate their results. All of the benchmarks they ran, plus several other application examples, are available publicly. You can download them from Graphcore's GitHub repository by clicking the button below.
Additionally, our own benchmarking team has created a quick start tutorial for those individuals that are looking to benchmark BERT on Graphcore IPUs. The tutorial is setup for those looking to use the Cirrascale Cloud Services platform for their training. You can download the Cirrascale Cloud Services BERT IPU Quick Start tutorial code by clicking the button below.
If you're interested in learning more about Graphcore IPUs, would like to use Graphcore IPUs on the Cirrascale Cloud Services platform, or would like to purchase a Dell EMC DSS 8440 machine intelligence server with Graphcore IPUs, visit us at http://graphcore.cirrascale.com.