Nvidia’s future challenger? Chinese start-up reveals aggressive AI chip road map
As US sanctions block access to advanced lithography, a Shanghai-based firm is betting on chip redesign and 3D memory stacking to break through bottlenecks
The Shanghai-based firm announced on Monday that its strategy was built on software-defined computing and 3D-stacked near-memory architecture, which it said could reduce reliance on the advanced manufacturing processes and cutting-edge memory currently restricted by Washington.
“We have to forge a path of our own,” founder Wei Shaojun said at the launch event. “That path cannot be about passively catching up within a framework set by others. We need independent architecture, original technology, a self-sustaining ecosystem and a secure, controllable supply chain.”
To that end, Dongfang Suanxin – or Shanghai Oriental Computing Technology – aims to circumvent traditional hardware constraints by changing how chips process data.
Software-defined computing reconfigures a chip’s computing and data-flow resources on the fly, tailoring the hardware to different workloads. Meanwhile, 3D-stacked near-memory architecture places memory layers vertically and closer to the computing cores rather than side by side on a circuit board, shortening the distance data must travel and reducing both processing delays and energy consumption.
The company debuted its flagship DF1000, a 14-nanometre AI processor reputed to deliver 520 teraflops of computing performance using the BF16 numerical format – a staple for training artificial intelligence models. Company specifications state that it also provides 6.4 terabytes per second of memory bandwidth and 900 gigabytes per second of scale-up bandwidth for inter-chip communication.
The chip was ready for mass production, with shipments expected to begin by the end of this year, according to the company.