NIPS 2018 Expo Demo

Dec. 2, 2018

Expo Schedule »

Performance Breakthrough with Habana’s AI Inference Processor

Sponsor: Habana Labs

Abstract:

We are bringing a pure-AI processor chip (manufactured by a deep submicron semiconductor process - not an FPGA) with full programmability which can run any neural network topology. The processor is in a PCIE line card that resides in an existing server and will be performing variety of general purpose inference workloads demonstrating the entire software stack. Including all the required steps from a high-level framework (e.g., MXNet/ONNX) all the way to a graph compiler recipe suited for the suggested hardware.

Users can activate the demos, such as type in text and see how the system predicts their sentiment, enter English text and see how it gets translated to German, see classification results of ResNet-50 inference for images selected by the user, See the performance of the image classification (ResNet50) by observing the time it takes the line-card to process 15,000 images and provide results back in the host memory, and interact with the demo team directly. By the demo date we may be adding additional typologies to the demo list. We would be bringing a laptop, small desktop computer.