【Hacker News搬运】谷歌首个张量处理单元：架构

hackernews

Title: Google's First Tensor Processing Unit: Architecture

谷歌首个张量处理单元：架构

Text:

Url: https://thechipletter.substack.com/p/googles-first-tpu-architecture

很抱歉，在使用 webscraper 抓取指定内容时遇到了连接超时的问题。由于无法连接到指定的 Substack 文章，我无法直接分析或总结其内容。如果您有其他请求或需要帮助，请告诉我。

Post by: c_joly

Comments:

hipadev23: How is it that Google invented the TPU and Google Research came up with the paper on LLM and NVDA and AI startup companies have captured ~100% of the value

hipadev23: 为什么谷歌发明了TPU，谷歌研究公司提出了关于LLM和NVDA的论文，人工智能初创公司已经获得了约100%的价值

nl: On the podcast interview now Groq CEO Jonathon Ross did[1] he talked about the creation of the original TPUs (which he built at Google). Apparently originally it was a FPGA he did in his 20% time because he sat near the team who was having inference speed issues.They got it working, then Jeff Dean did the math and the decided to do an ASIC.Now of course Google should spin off the TPU team as a separate company. It's the only credible competition NVidia has, and the software support is second only to NVidia.[1] <a href="https://open.spotify.com/episode/0V9kRgNS7Ds6zh3GjdXUAQ?si=qE3MLNuFQLCFmSnvhG87jQ&nd=1&dlsi=3752b373bdfd4f2c" rel="nofollow">https://open.spotify.com/episode/0V9kRgNS7Ds6zh3GjdXUAQ?si=q...</a>

nl: 在播客采访中，Groq首席执行官乔纳森·罗斯（Jonathon Ross）[1]谈到了最初的TPU（他在谷歌建立的）的创建。显然，最初他在20%的时间里做的是FPGA，因为他坐在有推理速度问题的团队附近 他们成功了，然后Jeff Dean计算了一下，决定做一个ASIC 现在，谷歌当然应该将TPU团队分拆为一家独立的公司。它；这是NVidia唯一可靠的竞争对手，软件支持仅次于NVidia [1]<a href=“https://；&#x2F；open.spotify.com/x2F；插曲&#x2F：0V9kRgNS7Ds6zh3GjdXUAQ？si=qE3MLNuFQLCFmSnvhG87jQ&amp；nd=1&amp；dlsi=3752b3bdfdf4f2c”rel=“nofollow”>https://&#x2F；open.spotify.com&#x2F；插曲&#x2F；0V9kRgNS7Ds6zh3GjdXUAQ？si=q</a>

formercoder: Googler here, if you haven’t looked at TPUs in a while check out the v5. They support PyTorch/JAX now, makes them much easier to use than TF only.

formercoder: 谷歌人在这里，如果你已经有一段时间没有看TPU了，看看v5。他们支持PyTorch；JAX现在使它们比仅使用TF更容易使用。

xrd: This article really connected a lot of abstract pieces together into how they flow through silicon. I really enjoyed seeing the simple CISC instructions and how they basically map on to LLM inference steps.

xrd: 这篇文章将许多抽象的部分连接在一起，了解它们是如何流过硅的。我真的很喜欢看到简单的CISC指令，以及它们如何基本上映射到LLM推理步骤。

kleton: Which ocean creature name is the current TPU?

kleton: 当前的TPU是哪个海洋生物名称？