Integrated Circuits for AI Data Centers
From Semiconductor Scaling to Photonic and Quantum Computing
Buch in deiner Nähe kaufen
Beschreibung
The data center has become the unit of computation for artificial intelligence, and the integrated circuit has become its limiting factor. AI models scale on one curve, while transistors, interconnects, memory, and power delivery scale on entirely different ones. The widening distance between them now decides what AI can afford to do. Most of the energy an AI data center consumes goes into moving data rather than performing arithmetic on it. That single fact reorders everything downstream: which architectures win, where the capital goes, and how much further CMOS can carry the industry. This book follows that logic from the transistor to the rack, beginning with how CPU, GPU, TPU, and custom accelerator architectures differ. It then traces the chip developments from Nvidia, AMD, Arm, and Intel now serving the AI ecosystem of OpenAI, Meta, Microsoft, Google, Apple, and xAI. Later chapters work through the crossover from copper to light that governs interconnect design, then rise to system-level architecture and data center economics. The closing chapters look past CMOS to hyperdimensional and neuromorphic computing, silicon photonics, advanced materials, and quantum processors, and ask which will run real workloads, and when. Written by leaders from industry and academia, this is a unified playbook for architects, system designers, product managers, and investors who need to translate semiconductor trends into AI performance, cost, and energy efficiency.
- Explains what separates CPU, GPU, TPU, and custom AI accelerators, and when each architecture wins
- Traces how large language models are mapped onto silicon, and what that mapping costs
- Shows where energy actually goes in AI data centers: moving data, not computing on it
- Maps the copper-to-optics crossover that now governs AI data center interconnect design
- Assesses photonic, neuromorphic, and quantum computing on when they will run real AI workloads
The data center has become the unit of computation for artificial intelligence, and the integrated circuit has become its limiting factor. AI models scale on one curve, while transistors, interconnects, memory, and power delivery scale on entirely different ones. The widening distance between them now decides what AI can afford to do. Most of the energy an AI data center consumes goes into moving data rather than performing arithmetic on it. That single fact reorders everything downstream: which architectures win, where the capital goes, and how much further CMOS can carry the industry. This book follows that logic from the transistor to the rack, beginning with how CPU, GPU, TPU, and custom accelerator architectures differ. It then traces the chip developments from Nvidia, AMD, Arm, and Intel now serving the AI ecosystem of OpenAI, Meta, Microsoft, Google, Apple, and xAI. Later chapters work through the crossover from copper to light that governs interconnect design, then rise to system-level architecture and data center economics. The closing chapters look past CMOS to hyperdimensional and neuromorphic computing, silicon photonics, advanced materials, and quantum processors, and ask which will run real workloads, and when. Written by leaders from industry and academia, this is a unified playbook for architects, system designers, product managers, and investors who need to translate semiconductor trends into AI performance, cost, and energy efficiency.
Explains what separates CPU, GPU, TPU, and custom AI accelerators, and when each architecture wins Traces how large language models are mapped onto silicon, and what that mapping costs Assesses photonic, neuromorphic, and quantum computing on when they will run real AI workloads Maps the copper-to-optics crossover that now governs AI data center interconnect design
Autor*in
Vikas Choudhary
Themen in »Integrated Circuits for AI Data Centers«
AI Hardware Design Data Centers Artificial intelligence Integrated Circuits for AI GPU TPU