China's analog AI chip could work 1,000 times faster than Nvidia GPU: study (www.scmp.com)

A team publishing in Nature Electronics on October 13 reports a Chinese analogue AI chip that, in their benchmarks, delivers roughly 1,000× higher throughput and about 100× better energy efficiency than state-of-the-art digital processors (e.g., Nvidia GPUs) while achieving accuracy comparable to digital systems. The researchers say they have addressed the historical "precision" bottleneck of analogue computing — which processes continuously varying values instead of binary zeros and ones — enabling performance and power gains without sacrificing model accuracy. If validated and manufacturable at scale, the work could be significant for AI/ML infrastructure: analogue devices promise much faster and more energy-efficient execution of compute-heavy tasks such as neural inference (and potentially parts of training), which would reshape datacenter economics and edge deployment. Important caveats remain: independent replication, production yield, integration with existing digital toolchains, programmability and generality across model types must be demonstrated before analogue chips can challenge GPU dominance. Still, the peer-reviewed claims mark a notable step toward practical analogue accelerators that trade binary precision for large gains in throughput and energy efficiency.