Domestic GPU performance leaps as Moore Threads S5000 reaches 60% of NVIDIA H100 in optimized AI tests
Moore Threads, a rising domestic GPU developer, recently went public with a valuation exceeding ¥400 billion (≈$55 billion USD). The company has now unveiled a new GPU architecture that significantly boosts both AI and gaming performance, with gaming workloads reportedly improving by up to 15x.
Even its current‑generation cards continue to receive engineering optimizations. SiliconFlow, one of Moore Threads' ecosystem partners, announced that it has achieved a major breakthrough in AI inference performance on the MTT S5000 GPU. After system‑level tuning and FP8 precision acceleration, the S5000 now delivers more than 4,000 tokens/s in prefill throughput and over 1,000 tokens/s in decode throughput.
To put this into perspective, NVIDIA's H100 achieves around 6,500 tokens/s in the same prefill scenario. With more than 4,000 tokens/s measured on the S5000, Moore Threads has reached over 61% of the H100's performance in this stage - a significant milestone for a domestic GPU.
The MTT S5000 is built on the company's Pinghu GPU architecture and supports FP8 for the first time, reaching up to 1,024 TFLOPS of FP8 compute. For comparison, NVIDIA's H100 delivers close to 4,000 TFLOPS FP8, meaning Moore Threads still trails in raw hardware capability. However, the latest results show that engineering optimization can meaningfully close the gap in real workloads.
Moore Threads acknowledges that domestic GPUs still face limitations in hardware scale, manufacturing process, and especially software ecosystem maturity. NVIDIA's CUDA ecosystem remains far ahead. But with increasing collaboration among Chinese developers and rapid improvements in both hardware and software, the company believes that domestic GPUs will continue to narrow the gap - and may eventually pressure NVIDIA in certain segments.
Moore Threads, founded in 2020 and headquartered in Beijing, is one of China's leading GPU and accelerated‑computing companies. The firm develops its own MUSA GPU architecture and produces graphics processors for AI, gaming, data‑center computing, and digital‑twin applications. Despite being a young company, Moore Threads has rapidly expanded its hardware and software ecosystem, aiming to build a fully domestic alternative to global GPU platforms. Its S‑series GPUs represent a major step toward strengthening China's independent computing‑hardware capabilities.
