谷歌/博通TPU硬核在哪里 | AVGO stock | AI 所谓泡沫不会破了但会巨量调整和套现

我的观点，是泡沫下一次破以后（两天跌出40%以上以后），我会毫不犹豫抄底买入博通（avgo）和K L A C(半导体设备龙头）。基本是S P的百倍股这样。买下来基本不需要管。爬升速度基本是sp一百倍。

博通是华人领导，broadcom 自2015为谷歌量身定做了tpu的核心元器件而且是代工logistics的高参。谷+博的路子和老黄完全不一样。博通的护城河比nvidia神。

现在博通是老博通和大部队avago合并以后，类似united和continental合并以后叫united，但是其实是两个大军合并。类似八路军和新四军整编成了P L A人民子弟兵。

 

说实话，台积电和老黄也就是和谷歌/博通打个平手。

S P： standard and then poor.

拒绝S P index

远离E T F。

本人拙见。

 

T P U和G P U本没有区别，都是做矩阵向量和张量运算的。类似英特尔C P U是算盘，老黄G P U是计算器，T P U出生就是速二代，类似一台电脑对计算器。输赢就不用想了。而且谷歌博通试图用光通讯概念（基本是不太可能，但是起码有说头）。

T P U出身就是高速计算，算了简单乘法就是降维业余玩一下。小手指一动，英特尔就完了。

 

Broadcom's primary role in Google's Tensor Processing Units (TPUs) is as the exclusive design partner and manufacturer for the custom application-specific integrated circuits (ASICs). While Google designs the core AI architecture, Broadcom is responsible for turning those blueprints into physical, manufacturable chips and managing the entire supply chain. [1, 2, 3, 4]

 

Key Contributions of Broadcom to the Google TPU

• ASIC Design and Manufacturing Execution: Broadcom translates Google's architectural specifications and intellectual property (IP) into a form that can be mass-produced at foundries like TSMC.
• Provision of Key Technologies: Broadcom provides critical, proprietary intellectual property (IP) and components, notably high-speed serializer-deserializer (SerDes) interfaces and networking solutions that enable the chips to communicate within Google's massive data centers and AI clusters.
• Supply Chain Management: Broadcom handles the complex logistics of the semiconductor supply chain, including securing access to advanced foundries, sourcing high-bandwidth memory (HBM), and managing the packaging, testing, and assembly of the chips into server boards.
• Risk and Cost Management: By partnering with Broadcom, Google offloads significant manufacturing and operational risks, allowing Google's internal teams to focus on core architectural innovations and software integration.
• Expertise and Scale: Broadcom's extensive experience in custom silicon and established relationships with suppliers ensure faster development cycles and cost-effectiveness that would be difficult for Google to replicate entirely in-house. [1, 3, 5, 6]

Broadcom has been the primary partner for every generation of Google's TPUs since their inception in 2015, and the collaboration continues to be a major source of revenue for Broadcom's custom AI chip business. [1, 2]

AI responses may include mistakes.

[1] https://www.theregister.com/2023/09/22/google_broadcom_tpus/

[2] https://www.techbuzz.ai/articles/broadcom-surges-10-as-google-s-ai-chip-partnership-pays-off

[3] https://forums.anandtech.com/threads/what-role-does-broadcom-play-in-semi-custom-asic-like-google-tpu.2624335/

[4] https://en.wikipedia.org/wiki/Broadcom

[5] https://newsletter.semianalysis.com/p/broadcoms-google-tpu-revenue-explosion

[6] https://en.wikipedia.org/wiki/Tensor_Processing_Unit

SerDes, which stands for Serializer/Deserializer, is a technology that converts parallel data into a serial bitstream for efficient transmission and then converts it back to parallel on the other end. It is a critical component in high-speed communication systems, enabling faster data transfer over fewer wires, reducing signal skew and noise, and allowing for higher bandwidth.

 

How it works

• Serializer: Takes multiple parallel data lines and converts them into a single, high-speed serial stream.
• Deserializer: Receives the serial stream and converts it back into the original parallel data format at the other end.

Common SerDes techniques

• Parallel clock: The serializer sends the parallel data and a separate clock signal. The deserializer uses this external clock to reconstruct the data.
• Embedded clock: The serializer embeds a clock signal within the serial data stream, eliminating the need for a separate clock wire. The deserializer then recovers the clock from the data stream.
• 8b/10b encoding: The serializer converts each 8-bit data byte into a 10-bit symbol to ensure better signal integrity by limiting the number of consecutive 0s or 1s. The deserializer decodes the symbols back to the original data.
• Bit-interleaved: The serializer multiplexes several slower serial streams into a single, faster stream by interleaving bits from each stream. The deserializer reverses this process.

Key applications

• Automotive: Simplifies the design of systems by transmitting high-definition video from cameras to electronic control units over a single cable.
• Data centers and AI infrastructure: Provides the high-speed links needed for connecting processors, memory, and other components for fast data transfers.
• Consumer electronics: Used in devices like smartphones, laptops, and gaming consoles to transmit data between components.

Why it's important

• Reduces complexity: Uses fewer wires for high-speed communication, which reduces the size, weight, and cost of cables and connectors.
• Increases speed: Enables much higher data transfer rates compared to traditional parallel interfaces, which is essential for modern high-bandwidth applications.
• Improves signal integrity: Reduces signal skew and crosstalk by transmitting data on a single line, leading to more reliable data transmission.

 

AI responses may include mistakes.