Chinese GPU maker Moore Threads unveiled its Lushan gaming and Huashan AI GPUs at MUSA 2025, promising massive performance leaps. Built on the new Flower Harbor (Huagang) architecture, Lushan delivers 15x AAA gaming performance, 50x ray tracing boost, and critically—full DirectX 12 Ultimate support, addressing its previous lineup’s major shortcoming.
Moore Threads Lushan & Huashan: Revolutionary Specifications
Performance Claims Comparison
| Metric | Lushan Gaming GPU | Huashan AI GPU |
|---|---|---|
| Architecture | Flower Harbor (Huagang) | Flower Harbor (Huagang) |
| AAA Gaming | 15x performance boost | N/A |
| Ray Tracing | 50x performance boost | N/A |
| AI Compute | 64x performance boost | Comparable to Blackwell B200 |
| Geometry Processing | 16x performance boost | N/A |
| Texture Fill Rate | 4x performance boost | N/A |
| Atomic Memory | 8x performance boost | N/A |
| Memory Capacity | 4x increase (up to 64GB) | 8 HBM sites, dual-chiplet |
| DirectX Support | DX12 Ultimate (first time) | N/A |
| Launch Timeline | 2026 | 2026 |
Architectural Breakthrough
The Flower Harbor architecture represents Moore Threads’ most ambitious design, addressing critical weaknesses that plagued previous MTT S80/S90 consumer GPUs. Full DirectX 12 Ultimate compliance enables modern gaming features including Mesh Shaders, Variable Rate Shading, Ray Tracing, and Sampler Feedback—all absent from earlier generations stuck at Feature Level 11_1.
Zhang Jianzhong, founder and CEO of Moore Threads (former 14-year Nvidia veteran and China general manager), emphasized the 50% increase in compute density and 10x energy efficiency improvement on the same process node. The architecture supports FP4 through FP64 mathematical precision formats end-to-end, plus proprietary mixed low-precision formats MTFP6 and MTFP4 for optimized AI workloads.
Lushan Gaming GPU: Targeting Global Competition
Lushan succeeds the MTT S80/S90 lineup with dramatic performance claims: 15x AAA gaming performance, 50x ray tracing throughput, 64x AI compute, 16x geometry processing, 4x texture fill rate, and 8x atomic memory access. With 4x memory capacity (current models feature 16GB GDDR6), Lushan could offer up to 64GB VRAM—surpassing consumer offerings from NVIDIA, AMD, and Intel.
The GPU integrates AI generative rendering (AGR) blocks under the UniTE unified rendering architecture, enabling mixed classic graphics pipelines with AI-assisted rendering paths in single workflows. The second-generation hardware ray tracing engine targets Neural Rendering and Path Tracing capabilities, positioning Lushan as China’s answer to RTX 5000/Radeon RX 9000 series.
Huashan AI GPU: Datacenter Ambitions
Huashan features dual-chiplet design with 8 HBM memory sites, targeting large-scale AI training clusters of 100,000+ GPUs. Moore Threads compares Huashan’s floating-point compute to NVIDIA Blackwell B200, with total bandwidth matching B200 while claiming superior memory access capacity and 4x memory capacity advantages.
The MTT S5000 GPU, designed to compete with NVIDIA Hopper, achieved 1,000 tokens/second in Decode and 4,000 tokens/second in Prefill on DeepSeek V3—respectable performance against established competitors. MTLink 4.0 interconnect plus Ethernet compatibility enables up to 1,314 GB/s inter-chip bandwidth with SHARP support.
Geopolitical and Market Context
Founded in 2020 by Zhang Jianzhong after leaving NVIDIA, Moore Threads went public on Shanghai Stock Exchange in December 2025, with shares surging five-fold. The company represents China’s “domestic substitution play” as US export restrictions force Chinese buyers toward local alternatives.
Analysts note valuations may reflect political momentum rather than technical fundamentals. “In the context of the US-China tech war, valuations have detached from reality, driven by politics rather than logic,” said Shen Meng of Chanson & Co. However, government-backed procurement mandates create guaranteed demand, positioning Moore Threads for profitability by 2027 despite current losses.
Launch Timeline and Competition
Both Lushan gaming GPUs and Huashan AI accelerators launch in 2026, marking Moore Threads’ most ambitious market entry. Without full specifications or benchmarks, direct comparisons to NVIDIA RTX 5000, AMD RDNA 4, or Intel Arc Battlemage remain speculative. However, full DX12 Ultimate support removes a major barrier to Western market acceptance.
For updates on Moore Threads GPUs and Chinese GPU competition, explore our coverage.
