Hardware Co-Design
FPGAsPermalink
TRIP (An Ultra-Low Latency, TeraOps/s Reconfigurable Inference Processor for Multi-Layer Perceptrons) is a “soft” deep neural network for MLPs. It features a hybrid OpenCL-Verilog design, where the MLP engine is written in Verilog and OpenCL is used for data exchange between the CPU and the FPGA. It runs on the Nallatech 385A platform (Arria10).
The table shows how such a targeted, custom design can outperform more general-purpose vendor offerings. We continue improving the design with features such as mixed precision, more efficient data movement, and a scalable OS/runtime. This project is a collaboration with Boston University.
Thermal-Aware ComputingPermalink
Cooling power is a non-negligible issue in HPC systems, and thermal
variations at every level (environment, hardware from transistors to
clusters) make it worse. The figure shows how running the same code on
four nodes of the same cluster can result in significantly different
temperatures of the compute elements. We are studying CPUs, GPUs, and
FPGAs and are using machine-learning prediction models for better task
scheduling on the system. This project is a collaboration with
Northwestern University.
Kazutomo Yoshii (ANL)