[News] China Claims New Architecture Using 14nm Logic and 18nm DRAM Could Match NVIDIA 4nm GPUs

Under tightening U.S. export restrictions, China is turning to innovative chip architectures to bolster computing capabilities. According to South China Morning Post, this strategy—pairing mature-node chips with novel system designs—aims to narrow the performance gap with NVIDIA.

South China Morning Post cites Wei Shaojun, vice-president of the China Semiconductor Industry Association, who argues that 14nm logic chips, when combined with high-performance memory and an advanced computing architecture, could approach the capabilities of NVIDIA’s 4nm AI-training processors. He further emphasizes that his proposed framework is built on a “fully home-grown supply chain.”

Wei’s approach centers on “software-defined near-memory computing,” using 3D hybrid bonding to stack 14nm logic dies with 18nm DRAM. Near-memory computing places processors closer to memory in a hardware–software co-design to reduce data movement and improve efficiency, the report adds.

Tom’s Hardware notes 3D hybrid bonding (copper-to-copper and oxide bonding), which replaces solder bumps with direct copper interconnects at sub-10 µm pitches, is a key element of the proposed design.

The solution also seeks to reduce market dependence on NVIDIA’s CUDA ecosystem, according to Commercial Times. Wei identifies CUDA as a major point of vulnerability, noting that once software, models, and hardware become tied to a single proprietary platform, deploying alternative processors becomes exceedingly difficult, as noted by Tom’s Hardware.

Chinese outlet PCPOP indicates that Wei has disclosed additional performance claims, stating that the domestic design achieves 2 TFLOPS per watt and a total of 120 TFLOPS—figures he said surpass NVIDIA’s A100 GPU. However, official specifications show that the A100 reaches up to 312 TFLOPS, well above the Chinese solution’s capabilities.

The underlying concept of using stacking design to offset limits in process technology is being echoed by other Chinese tech leaders. According to South China Morning Post, Huawei founder Ren Zhengfei argues that China can reach cutting-edge performance through chip “stacking and clustering,” without engaging in a direct node-by-node contest.

Read more

(Photo credit: FREEPIK)

Please note that this article cites information from South China Morning Post, Tom’s Hardware, Commercial Times, and PCPOP.