Graphical Processing Units are de-facto standard for acceleration of data parallel tasks in high performance computing. They are widely used to accelerate batch machine learning algorithms. High-end discrete GPUs are characterized by a very high number of cores (thousands), high bandwidth memory optimized for the stream access and high power requirements. Integrated GPUs are characterized by a medium number of cores (hundreds), medium bandwidth memory shared with CPU optimized for the random access and low power requirements. Data stream processing applications are often required to provide response within the limited time frame, operate on data in relatively small increments and have strict power requirements if deployed on the embedded devices. This work evaluates performance of integrated and discrete GPUs belonging to the same chip family on several variants of k-nearest neighbours algorithm over sliding window and stochastic gradient descent using OpenCL and novel Heterogeneous System Architecture platforms. We conclude that integrated GPUs provide a niche solution catering for to small work sizes that offers better power efficiency and simplicity of deployment.
展开▼
