Meta Platforms 近日宣布推出了其新一代人工智能（AI）训练与推理芯片，该芯片的性能相较于初代产品提升了三倍。这款名为MTIA的定制芯片系列，是Meta专为优化AI工作负载而设计的。目前，MTIA主要用于训练和优化排名及推荐算法，不过，Meta有着更远大的目标，即逐步扩展芯片功能，使其能够支持生成式人工智能的训练，例如其自家的Llama语言模型。
新一代MTIA芯片在设计上注重计算能力、内存带宽和内存容量的平衡，以提供更高效的AI运算。据透露，该芯片配备了256MB的片上内存，主频达到1.35GHz，这些规格显著超越了第一代的128MB内存和800MHz主频。这一提升得益于采用了台积电先进的5纳米工艺制造技术，确保了芯片在性能增强的同时，也能保持良好的能效比。
Meta的这一创新举措预示着其在AI硬件领域的持续投入和进步，旨在为自家的社交媒体平台和人工智能应用提供更强大的底层支持。通过不断提升硬件性能，Meta有望在人工智能的研发和应用上取得更多突破，进一步推动AI技术在社交网络中的创新应用。
英语如下：
News Title: “Meta Unveils New AI Chip: Double the Performance, Targeting Generative AI”
Keywords: Meta AI Chip, Performance Boost, MTIA Release
News Content: Meta Platforms recently announced the launch of its next-generation Artificial Intelligence (AI) training and inference chip, boasting a threefold increase in performance compared to its predecessor. The custom chip series, named MTIA, has been designed by Meta specifically to optimize AI workloads. Currently, MTIA is primarily utilized for training and enhancing ranking and recommendation algorithms, but Meta has ambitious plans to gradually expand its functionality to support the training of generative AI, such as its own Llama language model.
The new MTIA chip emphasizes a balance between computational power, memory bandwidth, and memory capacity to deliver more efficient AI operations. It is equipped with 256MB of on-chip memory and operates at a clock speed of 1.35GHz, surpassing the first generation’s 128MB memory and 800MHz clock speed. This improvement is made possible by leveraging TSMC’s advanced 5-nanometer manufacturing process, ensuring enhanced performance with a favorable power-to-performance ratio.
This innovative step by Meta signals its ongoing investment and progress in the AI hardware domain, aiming to provide stronger foundational support for its social media platforms and AI applications. By continually improving hardware performance, Meta aspires to make more breakthroughs in AI research and application, further pioneering innovative uses of AI technology within social networks.
【来源】https://www.ithome.com/0/761/105.htm
Views: 1