Amazon 与 Cerebras 合作：用超大芯片加速 AI 推理
This AI chip has over 50 TIMES the silicon area of a conventional GPU die. And Amazon is partnering with the company behind it to make AI respond faster. 👀 Amazon and Cerebras are teaming up to tackle two different parts of AI: processing your question and generating the response. By splitting the work between their chips, the goal is to make AI faster and more efficient. And as AI assistants take on more complex tasks, that speed could make a big difference! #AI #Amazon #Cerebras #Semiconductors #Tech
Amazon 与 Cerebras 宣布合作，将 AI 推理拆分为两段：Amazon 的 Trainium 负责处理用户提问的 prefill 阶段，Cerebras 负责逐字生成回答的 decode 阶段。Cerebras 芯片的硅面积超过传统 GPU 裸片 50 倍，把快速内存与计算核心放在同一颗芯片上，以缓解 decode 阶段的内存带宽瓶颈。双方希望借此让 AI 助手更快返回结果。
- 浏览
- 5256
- 点赞
- 291
- 评论
- 4
- 分享
- 9
- 收藏
- 36
数字来自平台，为最近一次抓取时的数据。
以下是视频口播的自动转写，可能有错字。
Amazon makes its own ai chip, so why partner with a company that makes one roughly the size of a dinner plate? because processing your question and generating the answer put different demands on the hardware.
so first up, an ai model processes your prompt, that's called prefill, and a lot of that work can happen in parallel. then comes decode, which is generating the response, one piece of text after another.
those steps repeatedly need more data from memory, and if it can't arrive fast enough, which is a huge challenge, powerful computing hardware ends up waiting, sara brass puts fast memory alongside computing cores on one enormous chip.
so amazon and sara brass have announced a system that divides the work, amazon's trainium handles prefill, while sara brass handles decode.
so for example, say you ask an ai assistant to build you a booking page for your business. it writes the code, runs a test, finds an error, and tries again.
you're waiting through each of those rounds, faster inference could shorten the model's part of that process, so you get to a working page sooner, that is the connection between this.
Enormous chip and something you actually use. How quickly an ai assistant can turn your request into a result, which is significant in timing.
Share this with someone who loves tech. Also, i'm a huge fan of sara brass. They're building some incredible stuff.
来源：TikTok · Tiff in Tech · tiktok.com