TECH – At a recent summit in Beijing, China Semiconductor Industry Association vice-chair Wei Shaojun claimed that China is developing a new AI accelerator using a mature 14 nm logic process and 18 nm DRAM — a chip design meant to rival Nvidia’s cutting-edge 4 nm-class GPUs.
According to Wei, the chip uses advanced 3D hybrid bonding to tightly stack memory and logic, dramatically boosting memory bandwidth and cutting down latency — a move aimed at overcoming the “memory wall” limiting many GPU-based AI tasks. He asserted the design could deliver a throughput of 120 TFLOPS while consuming just 60 watts, equating to about 2 TFLOPS per watt — a bold claim considering the slimmer fabrication node compared to Nvidia’s modern silicon.
This domestic design is part of a broader ambition to build a self-reliant AI hardware ecosystem within China, reducing dependency on Western suppliers and the CUDA GPU software ecosystem. The proposed architecture marries legacy chipmaking nodes with cutting-edge packaging — stacking 14 nm chiplets and 18 nm DRAM directly to enable near-memory computing, which promises lower energy consumption and faster data access.
Read More: Hangzhou Airport Deploys Robot to Repel Birds from Runway
Critics note this remains a theoretical design: no working silicon has been publicly demonstrated yet, and thermal dissipation, manufacturing yield, and real-world performance still present significant challenges. The dominance of CUDA and the entrenched software ecosystem built around Nvidia hardware also looms as a serious barrier for adoption of alternative architectures.
Nevertheless, this announcement reflects a strategic pivot — opting for architectural innovation and clever packaging rather than chasing ever-smaller node sizes. For a country facing export controls and supply-chain limitations, maximizing performance with older but proven processes could represent a pragmatic shortcut to narrowing the gap in AI hardware capability.
Whether this 14 nm-based AI chip will truly live up to its lofty claims — matching the grunt of 4 nm Nvidia GPUs — remains to be seen. But its ambition signals that China is betting heavily on hybrid-bonded silicon as a contender in the global AI hardware race.
