Architecture evolves AI models, says Blaize
Claimed to be the first architecture to enable concurrent execution of multiple neural networks and entire workflows on a single system, while supporting heterogeneous compute intensive workloads, the Graph Streaming Processor (GSP) architecture will be demonstrated at CES 2020 by Blaize.
The computing architecture offers advances in energy efficiency, flexibility, and usability, says the company for existing and new artificial intelligence (AI) in the automotive, smart vision, and enterprise computing segments.
The Blaize GSP architecture and Blaize Picasso software development platform blend dynamic data flow methods and graph computing models with fully programmable, proprietary SoCs. This allows Blaize computing platforms to exploit the native graph structure inherent in neural network workloads all the way through runtime, says the company. The massive efficiency multiplier is delivered via a data streaming mechanism, where non-computational data movement is minimised or eliminated for what Blaize claims is the lowest possible latency and it reduces both memory requirements and energy demand at the chip, board and system levels.
Blaize GSP is the first fully programmable processor architecture and software platform that is built from the ground up to be 100 per cent graph-native. All neural networks, whatever the types, are graphs, says Blaize. The inherent graph-native structure enables developers to build multiple neural networks and entire workflows on a single architecture, applicable to many markets and use cases. End-to-end applications can be built integrating non-neural network functions such as image signal processing with neural network functions, represented as graphs that are processed 10 to 100 times more efficiently than existing solutions, claims the company. The GSP allows AI application developers to build entire applications faster, optimise these for edge deployment constraints and run them efficiently using automated data-streaming methods, explains Blaize.
“Blaize’s vision of a native graph streaming processor (GSP) is relatively unique,” noted Karl Freund, senior AI analyst at Moor Insights & Strategy. “The GSP is more general purpose than, say, a single-function ASIC for AI, and can consequently create opportunities in many markets, from Automotive to the Edge to the Cloud.”
The Blaize architecture will debut at CES 2020 (7 to 10 January), LVCC, South Hall 2, Booth 25332.