Moore Threads details Huagang GPU architecture, with Lushan GPU claiming 15x AAA uplift and 50x ray tracing gains
Moore Threads used its MUSA Developer Conference 2025 stage to outline its next GPU architecture, Huagang , plus two chips built on it: Lushan for gaming and Huashan for AI. The company framed both as next year’s products, but it still held back full specifications and final SKUs.
Gaming upgrade: Lushan targets DX12 Ultimate, higher RT throughput, and more headroom for modern engines
For Lushan, Moore Threads is pitching a large jump versus its current consumer lineup. The slide shown at the event lists 15x “3A” game performance, 50x ray tracing performance, and 64x AI performance, plus 16x geometry processing, 4x texture fill, 8x atomic access, and 4x memory capacity.
Source: Moore Threads
The architectural work Moore Threads highlighted is also aligned with what its older cards struggled with: API feature coverage and RT capability. Huagang is presented as fully supporting DirectX 12 Ultimate , and it adds an AI generative rendering block (AGR) plus a second-generation hardware ray tracing engine.
Source: Moore Threads
Moore Threads also tied Lushan to its UniTE unified rendering design, with the aim of mixing classic graphics stages with AI-assisted rendering paths inside a single workflow.
Data-center upgrade: Huashan
Huashan is positioned as an “AI training + inference” chip aimed at large clusters. Moore Threads says Huagang brings a new instruction set, a claimed 50% increase in compute density, and a claimed 10x energy efficiency uplift on the same process, plus a new async programming model for scheduling and parallel execution.
On math formats, Moore Threads says Huagang supports FP4 through FP64 end-to-end, and adds its own mixed low-precision formats such as MTFP6 and MTFP4.
Comparison to NVIDIA Hopper and Blackwell, Source: Moore Threads
For scale-up, Chinese reports from the event say Huashan targets 100,000+ GPU “AI factories,” with MTLink 4.0 plus Ethernet compatibility, SHARP support, and up to 1314 GB/s inter-chip interconnect. Caixin also reports Moore Threads is talking about up to 1024 GPUs per node.
Moore Threads is also comparing Huashan’s floating point compute, memory bandwidth, and interconnect bandwidth as landing between NVIDIA Hopper and NVIDIA Blackwell, while claiming memory capacity better than both.
2026 might be quite interesting
Moore Threads is a GPU company, so it has to focus on data-center and AI, because it is publicly traded, and that is what investors tend to care about. Our focus, though, will be the gaming side, and we look forward to seeing MTT GPUs perform better than the current-generation series, which were fine for a newcomer but still well behind the competition (even if the pricing was reasonable).
Source: Sina Finance , Wccftech
