Moore Threads Pins Down Newer Generation Lushan GPU for Gaming, Huashan for AI GPUs
The Chinese graphics processing unit (GPU) manufacturer, Moore Threads, has now officially announced the next generation of its product portfolio, Lushan for Game and Huashan for AI applications. The revelation occurred during the MUSA 2025 conference, where Moore Threads explained its new "Flower Harbor" (Huagang) architecture, around which the company will build its market strategy this upcoming year.
Flower Harbor Architecture
The Flower Harbor architecture undergirds the new chips, which primarily incorporate the following major enhancements:
- Higher Density: Computes units newly redesigned at a 50% increase in density.
- Energy Efficiency: About 10% claims improvement in energy efficiency.
- Broad Compute Support: architecture ranges all forms of compute: FP64 to FP4, INT8, and proprietary MTFP formats.
- Scalability: For AI workloads, utilizes an asynchronous programming model and a high-speed MTLink interconnect, which allows connections of more than 100,000 GPUs.
Lushan: Leap in Gaming Performance
The new line of GPUs is intended to replace the old Lushan GPUs that served as the consumer MTT S80 and S90 graphics cards. Although full specifications are still under wraps, Moore Threads has set quite lofty goals for performance:
- Increase of up to 15x performance in modern games.
- Boost of 50x for ray tracing.
- 64x increase in performance in AI-related activities.
- 4x increase in video memory capacity.
Crucially, Moore Threads has announced full backing for modern graphics APIs like DirectX 12 Ultimate. This directly addresses one of the critical shortcomings of previous generation GPUs produced by the company. The architecture carries with it also a hardware ray tracing unit and UniTE, which is a unified system for neural rendering and path tracing.
Huashan: the AI and Data Center Targets
Geared toward data centers and AI-related tasks, the Huashan GPU is in a dual-chip solution featuring eight HBM memory stacks. Making direct comparisons with NVIDIA's Hopper and Blackwell architectures, Moore Threads claimed that Huashan is on par with the B200 in memory bandwidth but exceeds it in memory capacity and data access.
Ecosystem and Software Development
A major thrust of the announcement was developing a self-sufficient software ecosystem to rival platforms like NVIDIA's CUDA. Academician Zheng Weimin contended that the key is moving from chips that are just possible to use to ones that developers are "willing to use," pointing out the need for better development experience.
Responding to that, Moore Threads has enhanced its MUSA software architecture to version 5.0, compatible with programming languages like Triton and TileLang. The company is on the road to opening up most of its core components progressively to the developer community.
Availability
The first Lushan-based graphics cards and accelerators with the Huashan chip are expected to be made available in 2026.
