Moore Threads’ Quantum Leap:
15x Gaming Boost & 50x Faster Ray Tracing

The "Flower Harbor" Architecture Aims for the GPU Crown in 2026

At the MUSA 2025 Summit held on December 20, Chinese GPU manufacturer Moore Threads sent shockwaves through the tech industry. The company officially unveiled its next-generation "Flower Harbor" (Huagang) architecture, promising astronomical performance gains that could finally bridge the gap between domestic Chinese silicon and the global giants, NVIDIA and AMD.

1. Lushan: The New King of Chinese Gaming?

Named after the famous mountain, the Lushan GPU is Moore Threads’ latest consumer-focused chip. It is designed to replace the aging MTT S80 and S90 series with an unprecedented uplift in capabilities. Moore Threads is making some of the boldest claims in GPU history:

🎮 15x AAA Gaming Performance: A massive jump that could move Moore Threads from "entry-level" to "high-end enthusiast."
🔦 50x Ray Tracing Boost: Addressing the biggest weakness of previous MUSA chips with a completely redesigned RT engine.
🧠 64x AI Compute Increase: Positioning the card as a dual-threat for both gamers and local AI developers.
🛠️ DirectX 12 Ultimate Support: Finally achieving full parity with modern gaming APIs, including Mesh Shaders and Sampler Feedback.

2. Huashan: Challenging Blackwell in the Data Center

While Lushan targets gamers, the Huashan GPU is aimed squarely at the AI data center market. Built on the same Flower Harbor architecture, Huashan is a chiplet-based behemoth designed to compete with NVIDIA's Blackwell B200.

Early data suggests that Huashan offers Floating-Point compute performance and memory bandwidth that rivals the B200. More impressively, Moore Threads claims it actually exceeds Blackwell in memory access capacity. With MTLink high-speed interconnect, these GPUs can scale to clusters of over 100,000 units—a critical requirement for training the next generation of Large Language Models (LLMs).

Technical Specifications at a Glance:

The Flower Harbor architecture increases compute density by 50% and improves energy efficiency by 10%. It supports a wide array of formats, including FP64 down to the new MTFP4 mixed-low precision format, specifically optimized for high-speed AI inference.

3. 64GB VRAM: Ending the Memory Bottleneck

One of the most exciting details for enthusiasts is the 4x memory capacity increase. While the current MTT S80 features 16GB of GDDR6, Moore Threads teased that the new Lushan-based flagship cards could arrive with up to 64GB of VRAM. This would make it the highest-capacity consumer graphics card on the market, ideal for "Neural Rendering" and massive 8K texture packs.

4. Can They Deliver by 2026?

Moore Threads has set an ambitious target of 2026 for the launch of Lushan gaming cards. While the company has historically struggled with driver optimization, their recent progress—including the 1,000 Tokens/s performance of the MTT S5000—suggests they are moving faster than anyone expected. If they can pair this monstrous hardware with stable drivers, NVIDIA may finally face a "third-party" threat from the East.