An AI chip uses on-chip memory and a 14nm process to improve inference performance while reducing reliance on advanced semiconductor technologies.
Chinese startup Dongfang Suanxin (DFSX) has launched the DF1000, an AI chip that uses a 14nm process and custom memory integration to improve AI inference performance without relying on advanced manufacturing nodes or High Bandwidth Memory (HBM).
The chip places custom memory directly on the compute layer. This shortens the distance between memory and processing units, which can increase memory bandwidth and reduce data-transfer delays. The approach is intended to address some of the performance limits of using a 14nm process for AI workloads.
The design also reduces the chip’s dependence on advanced HBM, which Chinese chip companies have faced difficulty accessing because of US export restrictions. Instead, DFSX is using its chip architecture and memory arrangement to improve performance.
DFSX claims the DF1000 can deliver performance comparable to some mainstream Western chips for certain AI inference workloads. However, its training performance remains behind current high-end AI accelerators.
The chip can also reconfigure its hardware according to the software workload being processed. This allows its hardware resources to be adjusted for different AI applications instead of using the same configuration for every workload.
The DF1000 is not expected to compete with Nvidia’s latest AI accelerators across all workloads. Its importance lies in the approach taken to overcome restrictions on advanced semiconductor manufacturing and memory technologies.
DFSX has also outlined plans for two more generations of the chip. If the company can continue improving performance while using domestic manufacturing and alternative memory architectures, the approach could support China’s development of AI hardware without depending on the latest Western semiconductor technologies.
As AI features become more common in social networks, messaging platforms, cameras, and smart glasses, the industry will also need to address how public content, user identities, and AI-generated images are collected, processed, and used.
Click here for the original announcement.
