100 Graph-Processing Nodes in one UltraScale FPGA can replace 100 processor cores in DRC’s Massively Parallel Graph D

2016年1月21日 | By News | Filed in: News.

Source: https://forums.xilinx.com/t5/Xcell-Daily-Blog/100-Graph-Processing-Nodes-in-one-UltraScale-FPGA-can-replace/ba-p/677219

One of the demos in the Xilinx booth at SC15 late last year showed a massively parallel graph-processing application by DRC Computer Corporation based on a large number of graph-processing cores instantiated in one Xilinx Virtex UltraScale VU190 FPGA working in conjunction with an IBM POWER8 server to accelerate the algorithms by 100x. Graph processing using the Dijkstra and Betweenness Centrality algorithms used for analyzing social networks with large number of actors. Nodes can be people, places, or events and each node has attributes that describe relationships with other nodes in the network. These relationships are costly in terms of compute time and are therefore ideal for FPGA-based acceleration. Application for these algorithms include big data indexing and analytics, image processing, cryptography, data security, financial analysis, and fundamental research.

Frankly, as an old digital system designer, I had some trouble wrapping my mind around these social-network application concepts but a new 2-minute video of the demo presented by DRC COO Roy Graham, recorded at SC15, brought some clarity to the concepts for me:

Note: IBM has a strategic collaboration program with Xilinx and this DRC Computer Corporation demo is part of that program. You can read more about this topic in a new post on the SemiWiki site (“IBM’s OpenPOWER Presence Was Felt Heavily At SuperComputing ’15”) and in a previous Xcell Daily post (“IBM and Xilinx announce strategic collaboration to develop high-performance, energy-efficient cloud-computing apps”).


邮箱地址不会被公开。 必填项已用*标注