Scan to Download
Search "DeepSeek" has a total of4related results
Techub News reports that DeepSeek has exclusively published a detailed technical article on the Zhihu platform, systematically explaining its elastic computing platform, DeepSeek Elastic Compute (DSec), for the first time. This platform serves as the core training infrastructure for the DeepSeek V4.1 Agent model. The article elaborates on the architectural design of DSec, its resource scheduling strategies, and its technical advantages in supporting large-scale AI model training. This disclosure is seen as a concentrated demonstration of DeepSeek's AI infrastructure capabilities. (Qbit)
Techub News reports that AI company DeepSeek has announced the open-source release of its Ascend AI chip foundational components. This initiative aims to collaboratively build an efficient and user-friendly software ecosystem for Ascend AI chips with Ascend, lowering the barrier for developers to use the related hardware. The open-source components are expected to enhance software support and development efficiency for Ascend chips in AI training and inference tasks. (Quantum Bit)
Techhub News reports that NVIDIA has outlined three key elements for maximizing ROI in AI factories: productivity, durability, and interchangeability. Productivity refers to maximizing annual revenue potential by achieving the highest throughput per megawatt and the lowest cost per token; durability refers to NVIDIA GPUs and systems continuing to generate revenue years after shipment; interchangeability refers to their ability to run various AI and non-AI workloads, thereby broadening demand. The blog cites SemiAnalysis AgentX data, stating that NVIDIA's Vera Rubin NVL72 system achieves over 30 times higher throughput per megawatt compared to the GB300 NVL72, with up to 45 times lower cost per million tokens on the DeepSeek V4 Pro model. Additionally, NVIDIA's A100 GPUs, shipped since 2020, are still in commercial service, and CoreWeave has extended bookings for some of their units until 2029. (NVIDIA Blog)
Techub News, according to Decrypt, the U.S. NIST-affiliated agency CAISI released an evaluation report stating that DeepSeek V4 Pro lags behind U.S. frontier AI models by about 8 months. The agency used the IRT scoring system to evaluate based on nine benchmark tests, two of which are non-public datasets. This evaluation has raised questions among experts. The Stanford 2026 AI Index shows that the AI performance gap between China and the U.S. has narrowed to 2.7%, and DeepSeek's performance in public benchmark tests is close to that of top U.S. models. In addition, the cost comparison excluded most U.S. models, retaining only GPT-5.4 mini for comparison.
Current search results fully loaded
