Efficient Computer raises $97 million to scale dataflow-based low-energy chips
Low-energy chip startup Efficient Computer closed a $97 million round led by TQ Ventures, its second raise this year, lifting total funding to $650 million as it ramps shipments of its dataflow processor.
Eclipse, Union Square Ventures, Giant Ventures, Triatomic Capital, TO Capital, TF Capital, Mana Ventures, Toyota Ventures, Overmatch and Borderless also participated in the round. Efficient closed a $60 million round in February.
The company is developing a chip built on a dataflow architecture that it says can cut energy consumption by as much as 100 times compared with the x86 architecture used by Intel Corp. and Advanced Micro Devices Inc. According to Efficient, its chips reduce consumption to the point that they could in theory power devices for months or even years at a time.
Modern central processing units and graphics processing units prioritize high performance, low latency and massive throughput over power conservation, and as a result are limited by architectural overheads that include complex control logic, high-speed data movement and the need to maintain a precise execution state.
In a blog post announcing the round, co-founder and Chief Executive Brandon Lucia described the specialized processors used to accelerate AI workloads today as a "devil's bargain," because they give up programmability and adaptability to gain efficiency and speed, leaving them useful for only a narrow subset of tasks such as AI inference.
Efficient says its dataflow architecture removes the unnecessary data movement and architectural overheads intrinsic to current CPU and GPU designs. By distributing workloads and connecting instructions to reflect the application dataflow, it aims to achieve large gains in performance per watt without sacrificing performance or programmability.
Dataflow chips have been discussed in academic literature for decades but have seen little commercial traction because of the difficulty of programming them for the wide variety of tasks handled by chips from Intel, AMD and Nvidia. Interest in the approach has revived as demand grows for more energy-efficient AI processors.
Lucia said Efficient took the basic concepts of the dataflow architecture and redesigned both the chips and the software tools needed to make them as adaptable as general-purpose processors. The company ultimately wants to build larger chips for data center servers, but its first market is smaller: battery-powered robots and autonomous drones. "We sort of thread the needle where we're easy to program, fast and efficient," Lucia told Reuters. "When you build an AI system, the system ends up doing a lot more than two little nano-optimized AI algorithms."
Efficient has not disclosed the names of its customers or details of its revenue, but Lucia said the company is seeing "overwhelming demand" for its first product, the Electron E1 chip. The new funds will be used to scale production volumes and increase shipments to customers.
TQ Ventures partner Andrew Marks said the need for greater power efficiency extends beyond running AI models. "What convinced was Efficient's ability to build both the hardware and the software, and turn that breakthrough into a business," he said. "Not only have they taped out four times, but they're already shipping chips to customers at volume."