Meta发布新一代AI训练与推理芯片，性能大幅提升至初代三倍
Meta Platforms日前宣布推出其最新研发的AI训练与推理加速器项目（MTIA）的新一代芯片。据悉，此次发布的芯片性能表现显著，远超初代产品三倍。
MTIA是Meta专为AI工作负载设计的定制芯片系列。目前，该芯片主要应用于排名和推荐算法的AI训练。然而，Meta表示其长远目标是扩展芯片功能，开始训练生成式人工智能，如公司的Llama语言模型等。
新一代MTIA芯片注重在计算、内存带宽和内存容量之间实现适当的平衡，这被视为是一次“根本性”的突破。该芯片拥有256MB的片上内存和高达1.35GHz的主频，采用先进的台积电5nm工艺制造。相较于第一代产品的128MB内存和800MHz主频，这一进步极为显著。
此次发布的芯片将有望为Meta在AI领域的发展提供强大的技术支持，并可能对整个科技行业产生深远影响。未来，我们期待看到更多关于MTIA的进展，以及其在AI领域的广泛应用。
英语如下：
News Title: “Meta Unveils Super-Powered AI Training and Inference Chip: Performance Surge Ushers in a New Era of Generative AI”
Keywords: Meta releases new generation chip; three-fold performance boost; AI training and inference chip; MTIA chip upgrade
News Content:
Meta Platforms recently announced the release of its newly developed AI training and inference accelerator project (MTIA) next-generation chip. It is reported that the performance of this chip exceeds its initial generation by three times.
MTIA is a series of custom chips designed by Meta for AI workloads. Currently, the chip is mainly used for AI training in ranking and recommendation algorithms. However, Meta has stated that its long-term goal is to expand the functionality of the chip to begin training generative artificial intelligence, such as the company’s Llama language model.
The new generation MTIA chip focuses on achieving an appropriate balance between computation, memory bandwidth, and memory capacity, which is seen as a “fundamental” breakthrough. With 256MB of on-chip memory and a main frequency of up to 1.35GHz, it is manufactured using advanced TSMC 5nm process technology, which is a remarkable improvement from the first generation’s 128MB memory and 800MHz main frequency.
This released chip is expected to provide powerful technical support for Meta’s development in the field of AI and may have far-reaching impacts on the entire technology industry. In the future, we look forward to seeing more progress about MTIA and its widespread application in the AI field.
【来源】https://www.ithome.com/0/761/105.htm
Views: 1