Moore Threads MTT S5000 AI GPU Passes China Security Certification

Moore Threads has released the MTT S5000 AI accelerator, a high-performance GPU built on its PH100 chip. The company recently announced that this model passed China's national 'Reliable and Secure Evaluation', marking the first time an AI training or inference chip received this certification.

First AI training chip to receive national security certification in China.

The MTT S5000 utilizes Moore Threads' fourth-generation MUSA architecture, codenamed 'Pinghu'. This platform includes dedicated TCE, TME, and ACE engines designed to optimize data processing. The hardware supports full precision ranging from FP8 to FP64 and provides native adaptation for major AI frameworks like PyTorch, Megatron-LM, vLLM, and SGLang.

In terms of raw compute power, the single-card dense AI performance reaches 1000 TFLOPS. The device is equipped with 80GB of VRAM and delivers a memory bandwidth of 1.6TB/s. For multi-GPU setups, the inter-card interconnect bandwidth operates at 784GB/s to facilitate efficient data transfer between units.

Moore Threads claims that the MTT S5000 breaks through traditional architecture bottlenecks in large-scale computing. The company states this results in a significant end-to-end performance leap for massive training clusters and high-concurrency online inference services. They further assert that its stability matches international mainstream flagship products.

This certification is restricted to the Chinese market due to national security requirements. This certification underscores the increasing focus on secure and reliable domestic computing solutions in China. No pricing or availability details have been disclosed by the vendor at this time.