Ali PG1 Server 1536GB

The Ali AP1 (Alibaba PG1) is an ultra-dense 10U AI super-node engineered for the most demanding training and inference of LLMs, multimodal, and generative models.

This powerhouse integrates 16 Pingtou Ge Zhenwu 810E accelerators, providing a 1536GB of total HBM2e memory and an aggregate FP16 performance of 1968 TFLOPS.
With a staggering memory bandwidth of 2765 GB/s and a proprietary 700 GB/s inter-chip interconnect (ICN), the system achieves near-linear scaling for models with billions of parameters.

The host is equipped with dual 5th Gen Intel® Xeon® 8558P processors (96 cores total) and 2TB of high-speed ECC DDR5 RAM. Its networking is designed for hyperscale deployment, featuring ten 200Gbps Ethernet ports and dual 25Gbps links.
Fully optimized for DeepSeek and Qwen ecosystems and compatible with PyTorch/TensorFlow.

Application scenarios

Training the largest LLMs and multimodal models.
Large-scale AI clusters.
Generative AI and complex analytics.