Chinese GPU manufacturer Moore Threads has reportedly integrated DeepSeek-R1, an open-source model, into its proprietary graphics hardware. This development follows similar moves from NVIDIA, Microsoft, and AMD, which have explored various applications of DeepSeek-R1 since late January. Unlike these Western firms, Moore Threads operates in a market where its GPUs are still catching up in terms of performance.
Evaluations from early 2024 indicated that the company’s MTT S80 struggled to compete even with AMD’s integrated Radeon 760M solution. However, the emergence of DeepSeek’s models has sparked interest due to their relatively modest hardware requirements compared to the traditionally high computational demands of cloud-based processing. Tom’s Hardware has noted instances of open-source models running efficiently on cost-effective devices, including Raspberry Pi units.
Recent Chinese media reports indicate that Moore Threads has successfully deployed DeepSeek-R1-Distill-Qwen-7B on its MTT S80 desktop GPU. Additionally, the firm has extended support to its MTT S4000, a data center-oriented graphics card. According to an official statement, Moore Threads leveraged the Ollama open-source framework to optimize DeepSeek-R1-Distill-Qwen-7B for its hardware, claiming that the adaptation demonstrated robust performance across various Chinese-language tasks. The company emphasized that this effort underscores the flexibility and CUDA compatibility of its in-house GPU architecture.
Although Moore Threads has not disclosed detailed performance benchmarks or technical specifications, its announcement suggests an intent to position its GPUs as viable solutions for AI inference workloads. ITHome reported that users could deploy the DeepSeek-R1 model on both the MTT S80 and MTT S4000, with some users having already completed similar configurations manually.
Moore Threads highlighted its proprietary high-performance inference engine as a key factor in optimizing computational efficiency and resource management. The company claims that its software and hardware co-optimization strategy enhances DeepSeek-R1’s performance through customized operator acceleration and advanced memory management techniques. This, in turn, is expected to facilitate the deployment of more complex AI models in the future.
While the exact capabilities of Moore Threads’ GPUs remain uncertain compared to established industry leaders, the firm’s engagement with DeepSeek-R1 suggests an effort to carve out a niche in the growing demand for local AI inference solutions. The extent to which its hardware can compete with Western counterparts in real-world AI applications will depend on further testing and independent performance assessments.
Sources: ITHome, Tom’s Hardware, IT Bear China
