MTT S5000

MTT S5000 is a Universal GPU engineered for the generative AI era, purpose-built for large model training, inference, and high-performance computing. With Moore Threads' next-generation PH100 chip and the advanced PingHu architecture, it delivers full-precision compute from FP8 to FP64. Powered by the fourth-generation MUSA full-stack platform with native support for PyTorch, Megatron-LM, vLLM and SGLang, it enables near-zero-effort code migration. PH100 was among the first chips to receive security and reliability certification from the China Information Technology Security Evaluation Center...

Key Highlights

• Hardware-native FP8 with over 30% training improvement on DeepSeek/Qwen-class architectures
• designed for 10K-GPU clusters with up to 95% scaling linearity and chip-to-system RAS raising training success rate by over 30%
• per-card Prefill over 4000 Tokens/s and Decode over 1000 Tokens/s (PD-disaggregated, six 8-GPU S5000 servers, native FP8)
• 16 GPUs
• native FP64
• CUDA-compatible MUSA migration
• OAM modules in liquid-cooled and air-cooled versions.

Technical Specifications

Typical Use Cases

英文取自官网英文页；官网未公开 S5000 自身显存容量/带宽与单卡算力量化规格，页脚①②③为竞品对比基准、未采信为产品参数；官网 cover.jpg/banner.jpg 为纯背景空图（无产品主体）未采用，图片采用官网 row-3-1.jpg（MTT S5000 OAM 计算模组风冷版渲染图，含官方图注）。