据 Bloomberg 报道,DeepSeek 计划在内蒙古自建的大型数据中心部署至少 16 万颗华为昇腾 950DT 芯片,用于运行其模型,可能成为已知规模最大的华为 AI 芯片集群之一。
DeepSeek reportedly plans to deploy at least 160,000 Huawei Ascend 950DT chips in Inner Mongolia, potentially creating one of the largest known Huawei AI clusters!
Bloomberg reports the chips would run DeepSeek’s models. The company currently has no plans to use them for training, where it has so far relied on Nvidia.
How capable are they? Huawei’s announced specs include 144GB of high-bandwidth memory and 4TB/s of memory bandwidth. Nvidia’s H200 offers 141GB and 4.8TB/s. That puts Huawei close on memory capacity, with lower bandwidth. It does hwever not establish equivalent real-world performance.
Bloomberg describes the 950DT as broadly comparable to Nvidia’s older Hopper generation. Crazy at which scale china is building their national chip industry.
来源:@kimmonismus · x.com