查看原文
其他

Sora爆火后,如何解决算力焦虑?

                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    


近期

多家券商机构发布研报表示

Sora大模型的发布

将进一步拉大算力缺口


业内也认为

Sora将快速推进应用端商业化落地

拉动产业规模

进而传导到上游算力基础设施

算力需求将再次迎来爆发


点击图片,阅读更多有关Sora内容


Sora拉大算力缺口


算力是AI三大核心要素之一

重要程度不言而喻


以Sora为例

其延续了此前GPT模型

“源数据-Transformer-Diffusion-涌现”的技术路径

意味着其发展成熟

同样需要海量数据+大规模参数+大算力作为基座

视频训练所需数据量

远大于文本训练的数据量


经中信证券简单估算

对于一个60帧的视频(约6至8秒)

Sora要生成至少

约120万个Token(文本处理最小单位)

这是相当大的计算量


OpenAI的技术文档也透露

随着训练数据的增加

样本质量明显提升

这也意味着需要更多的算力来支撑



算力短缺成AI时代“难题”


算力短缺问题并非出现在近期

自ChatGPT推出以来

AI算力一直是挥之不去的阴影


在2023年4月

ChatGPT的付费项目就曾被迫暂停

这种情况凸显了一个不争的事实

在当前技术发展的大背景下

如何高效、可持续地提供充足的算力资源

成为全球科技企业共同面临的难题



Sora发布前

OpenAI的CEO萨姆·奥特曼表示

正在与投资者进行洽谈

旨在提高全球芯片产能

扩大其支撑AI性能的能力

该项目可能需要筹集

至多5万亿至7万亿美元资金


算力的稀缺和昂贵

已经成为制约AI发展的核心因素


它不仅仅是

让AI模型跑得慢点那么简单

还直接关系到

AI技术的商业化和产业化进程

没有足够的算力支持

再炫酷的AI应用

也只能停留在实验室里

无法走进寻常百姓家



伏羲智库创始人、

清华大学互联网治理研究中心主任

李晓东曾表示

算力提升和成本降低

可以让通用人工智能

来到普通百姓的身边


如何补齐算力“缺口”?


当下

全社会各领域对于算力的共性需求

呈现爆发式增长


预计未来五年

全球算力总规模的年均增速

将超过50%


2023年世界互联网大会乌镇峰会期间

中国移动董事长杨杰表示

面对这一趋势

要加快算力网络的产业实践

强化通算、智算、超算等多样算力统筹运用和大规模供给

推动算力成为像水电一样

可“一点接入、即取即用”的社会级服务


蚂蚁集团首席技术官何征宇呼吁

行业加强软硬协同与生态建设

共同提升计算基础设施

算力规模、可用性与宽容度

…….



2月27日

世界互联网大会

将在2024年世界移动通信大会期间举办

“算力网络”专题论坛

广邀全球业内大咖共议

“算力网络:智能网络赋能智慧世界”

就"网络助力高质量算力供给"

“应对AI计算需求的快速发展”

“打造低成本、绿色、可持续的算力”

三大议题分别展开讨论

共促变革时代的数智化转型





点击文末“阅读原文”进入世界互联网大会官网





相关阅读

脑机接口:治疗病人or制造“超人”?

大模型Sora刷屏!最新动态在这里

AI冒充“总统”来电,你能识别吗?

参考丨新华财经、北京日报客户端、每日经济新闻撰文:刘昊 编辑:李飞 排版:李汶键 统筹:李政葳

扫描二维码

关注我们


戳我~  一键进入官网

继续滑动看下一个
向上滑动看下一个

您可能也对以下帖子感兴趣

文章有问题?点此查看未经处理的缓存