Key Takeaways
1. Moore Threads introduced the Huagang architecture at the MUSA Developer Conference, aimed at gaming and AI, with a launch scheduled for next year.
2. The new “Lushan” GPU promises up to 15 times better performance for AAA games and 50 times for ray tracing, with claims of a second-generation hardware ray tracing engine.
3. Lushan is expected to have 64 GB of memory, up from 16 GB, and improvements in AI compute performance and other processing capabilities, although evidence for these claims is lacking.
4. The Huashan AI GPU may feature a dual-chiplet design, 9 HBM modules, and performance metrics that rival Nvidia’s offerings, with a significant expected boost in compute density and efficiency.
5. Although no gaming demos were available, a performance demo of the MTT S5000 GPU showed promising results, highlighting China’s push for GPU independence amid export restrictions.
Chinese GPU manufacturer Moore Threads recently hosted a MUSA Developer Conference, where they introduced their next-generation “Huagang” (also known as “Flowerpot”) architecture. This new architecture is scheduled to debut next year and aims to cater to both gaming and artificial intelligence needs. However, the conference did not provide many specific technical details about the architecture, focusing instead on bold performance claims.
Exciting New GPU Models
A new gaming GPU named “Lushan” is set to be developed using the Huagang architecture, taking over from the current MTT S80 and S90 models. The company boasts an incredible 15 times improvement in performance for AAA game rendering and a staggering 50 times enhancement in ray tracing capabilities. The upcoming GPU is also said to include a second-generation hardware ray tracing engine and full support for DirectX 12 Ultimate, which should improve compatibility. But, it’s crucial to note that there is no solid evidence yet to back up these claims, so skepticism is advised.
Memory and Performance Upgrades
In terms of memory, the Lushan GPU is expected to provide up to 64 GB of memory, an increase from the current 16 GB GDDR6 found in existing models. Moore Threads also asserts improvements in AI compute performance by 64 times, 16 times in geometry processing, 4 times in texture fill performance, and 8 times in atomic memory access. Furthermore, the GPU is rumored to have a new “UniTE” unified rendering architecture, complete with a dedicated AI hardware block. Yet, it’s still uncertain whether these assertions will prove to be true.
Teasing the AI GPU
Alongside the Lushan, there are hints about the Huashan AI GPU, which reportedly features a dual-chiplet design and 9 HBM modules. The company claims its performance will rival Nvidia’s Hopper and Blackwell GPUs, while memory bandwidth is said to surpass Nvidia’s B200. The AI GPU is also expected to support FP4 through FP64 computing with proprietary formats (MTFP4, MTFP6, MTFP8), and it could scale beyond 100,000 GPUs using the MTLink 4.0 interconnect, hitting 1314 GB/s. Moore Threads claims a 50 percent boost in compute density and a tenfold improvement in efficiency compared to current models.
Although there are no gaming demos ready for these new GPUs, a performance demo for the DeepSeek V3 on the MTT S5000 (which is another GPU set to release next year but is not part of the Huashan series) was shown. This GPU reportedly managed to achieve 1000 tokens/second in Decode and 4000 tokens/second in Prefill, indicating a slight edge over Nvidia’s Hopper lineup. The forthcoming GPUs symbolize China’s commitment to GPU independence in light of export restrictions, and further details are anticipated in the upcoming months as the launch approaches.
Source:
Link