In an era where artificial intelligence is progressing at a breakneck speed, Huawei has emerged as a significant player in the field, garnering considerable attention for its innovationsThe recent unveiling of an important AI patent from Huawei marks a crucial milestone that could reshape the landscape of AI technologies.

The patent titled "Quantization Method, Processing System and Quantization Unit for Artificial Intelligence Models" has recently been disclosed by Huawei Technologies Co., Ltd., and its implications are profound within the AI sectorAs AI technologies evolve, the size of models grows exponentially, leading to challenges related to large parameter sizes, memory consumption, and computational complexityHuawei's new patent addresses these critical issues that the industry grapples with.

According to the patent abstract, the quantization unit innovatively utilizes the processing granularity of various computing units within devices running AI models to segment quantization matrices into several adaptive block quantization matricesSubsequently, the quantization unit processes these segmented matrices, yielding quantized block matrices that alleviate the memory demands of AI modelsUnlike traditional methods, this quantization unit does not require outlier statistics at the tensor level, allowing real-time quantization of weight matrices and activation matricesSuch innovation significantly enhances the operational efficiency of AI models, tackling limitations of traditional methods regarding high computational complexity and a lack of real-time performanceFor real-world applications, this advancement means AI models can function smoothly on a broader range of devices, including smartphones and IoT gadgets, delivering quicker response times and a more seamless interaction experience for users.

It's worth noting that Huawei has disclosed a remarkable total of 26 patents in the AI domain since the start of the year

Advertisements

On February 5, at the Guangdong High-Quality Development Conference, Huawei's chairman, Liang Hua, emphasized that AI technology is at an accelerated development stage, entering a pivotal period for applications across various industriesInnovation, he stated, is the key to seizing new opportunitiesThe announcement of this patent exemplifies Huawei's unwavering commitment to innovation in the AI space.

Furthermore, collaboration with DeepSeek is deepening Huawei's foothold in the AI ecosystem by injecting substantial energy into domestic AI developmentOn February 6, Huawei China announced that its DCS AI full-stack solution, featuring the ModelEngine product, now fully supports the local deployment and optimization of DeepSeek's large models R1 & V3, along with its distillation series modelsThis marks a significant step towards accelerating customer AI applications to fruition.

The ModelEngine serves as a robust AI platform for training, inference, and application development of large models, offering a one-stop solution for users from data preprocessing to model training, optimization, and deploymentUsers can leverage the efficient inference capabilities of DeepSeek, with developers now able to achieve 'one-click deployment' of the DeepSeek series models, thus eliminating intricate steps like hardware selection, compatibility validation, and environment setupThis development greatly reduces the barriers for enterprises wishing to incorporate AI into their operations.

As the collaboration progresses, numerous achievements have emergedOn February 1, Huawei Cloud introduced its pioneering DeepSeek R1/V3 inference service based on Huawei's Ascend Cloud ServicesNotably, on February 4, models such as DeepSeek-R1, DeepSeek-V3, DeepSeek-V2, and Janus-Pro were officially made available on the Ascend community, allowing for quick deployment on Ascend hardware platforms through straightforward accessAdditionally, on February 5, the Xiao Yi assistant app, built on the native HarmonyOS, integrated DeepSeek-R1's beta version

Advertisements

Advertisements

Advertisements

Advertisements