There’s booming demand for silicon custom-designed to accelerate AI workloads, as the gobs of cash raised by startups like Hailo Technologies, Graphcore, and Untether AI demonstrates. The fierce competition isn’t deterring Blaize (formerly Thinci), which hopes to stand out from the crowd with a novel graph streaming architecture. The nine-year-old startup’s claimed system-on-chip performance is impressive, to be fair, which is likely why it’s raised nearly $100 million from investors including automotive component maker Denso.
Blaize emerged from stealth today with $87 million raised over several venture rounds from strategic and venture backers Denso, Daimler, SPARX Group, Magna, Samsung Catalyst Fund, Temasek, GGV Capital, SGInnovate, and Magna; the second-to-last round closed in September 2018 and totaled $65 million. The company initially focused on what it called vision processors — chips to speed up vision, radar, and sensor fusion tasks — before expanding to encompass datacenters, edge infrastructure devices, and enterprise client devices.
Blaize claims its chips are some of the first to enable concurrent execution of multiple AI models and workflows on a single system, while supporting a range of heterogeneous compute-intensive workloads. They’re fully programmable, with a streamlined processing pipeline that yields between a 10 to a 100 times improvement in raw performance, latency, and energy efficiency at the board and system level in addition to reduced size and cost.
“Blaize was founded on a vision of a better way to compute the workloads of the future by rethinking the fundamental software and processor architecture,” said Blaize cofounder and CEO Dinakar Munagala. “We see demand from customers across markets for new computing solutions that address the immediate unmet needs for technology built for the emerging age of AI, and solutions that overcome the limitations of power, complexity and cost of legacy computing.”
How does Blaize achieve the speedup? With “100% graph-native” hardware. Neural networks, a particular kind of machine learning algorithm comprising interconnected units called nodes (a collection of nodes is called a layer), receive numerical inputs and multiply those inputs by a value (weight) before passing them onto an activation function, which defines the node output. Graphs in the context of graph theory also comprise nodes (also called vertices or points), which are connected by edges or links (or lines).
With a graph-native structure like that of Blaize’s chips, which leverages graph computing models and a dynamic streaming mechanism that minimizes non-computational data movement, developers can build multiple neural networks and workflows on a single architecture through to runtime. Better yet, in tandem with Blaize’s Picassoä software, the company said networks can be built that integrate non-neural network functions such as image signal processing, all represented as graphs.
Blaize says its products are being used in datacenter servers, edge infrastructure platforms, and client platforms in business and consumer apps ranging from massive machine learning farms to sensor fusion and advanced neural networks for autonomous driving.
“The coming out of Blaize and its leading Graph Streaming Processor is extremely exciting,” said angel investor and former Intel executive vice president David Perlmutter. “As an initial investor in Blaize, I recognized early on the great efficiency of one of the first to market a complete solution designed from scratch, fully optimized for AI and Neural Network applications. The unprecedented efficiency is great for a wide range of edge applications, particularly the automotive market. I am proud of the team in delivering on the promise.”
In addition to its headquarters in El Dorado Hills, California, Blaize has teams in Campbell, California and Cary, North Carolina, as well as subsidiaries in Hyderabad, India and Leeds and Kings Langley in the U.K. It employs more than 325 people worldwide.