Moore Threads just pulled the wraps off its next-gen Flower Harbor architecture — and the numbers are brutal for US chip dominance.
15× gaming performance. 50× ray tracing. 64× AI compute. 4× memory capacity.
Their Lushan gaming GPU jumps straight into modern territory with full DirectX 12 Ultimate, AI-driven rendering, and next-gen ray tracing — areas US firms claimed China would never master. Meanwhile, the Huashan AI GPU is being benchmarked directly against NVIDIA Hopper and Blackwell, matching bandwidth, pushing memory access beyondB200, and scaling to 100,000-GPU clusters.
Even today’s Moore Threads S5000 is already hitting 1,000 tokens/sec decode and 4,000 tokens/sec prefill on DeepSeek models — squarely in Hopper territory.
Washington bet export controls would freeze China in time. Instead, Chinese companies rebuilt the stack, rewrote the architecture, and are now sprinting.
Expect more handheld PCs with homegrown Linux distro preinstalled in the future.