Skip to content
EN
Industry

Advanced Computing, Hardware & Semiconductors

AI drives the next generation of chips, computing architectures, quantum technologies, and scalable hardware systems

Image: Jakub Pabis · Pexels-Lizenz

Overview

AI Chips and Advanced Computing: Why Hardware Is Being Rethought

AI systems need more and more computing power, and shrinking conventional processors can no longer supply it on its own. Chips are now the base hardware layer of modern AI systems. [1, 9]

Why conventional scaling is reaching its limits

A 'Post-Moore's Era' is described, in which the doubling of transistor counts and speeds every two years no longer keeps up with AI workloads. Engineers face growing bottlenecks in heat dissipation and in communication on the chip. Making transistors smaller no longer produces the big jumps in performance people expect. It attributes this to limits set by physics, the difficulty of managing heat, and costs that keep climbing. It concludes that progress now depends on new architectures as well as miniaturization. A Nature special collection makes a similar point: gains from pure technology scaling of conventional processors cannot keep pace with AI's demands. [1, 5]

Specialized chips for AI workloads

AI work differs from classical computing: it relies on massive parallel processing, operations centered on data, and a mix of different tasks. Several chip types address this. GPUs were first designed for rendering graphics. They are good at parallel processing, which suits the training of AI models. TPUs were developed by Google for machine learning and speed up both training and inference of deep learning models. FPGAs and ASICs can be customized and are tuned for specific uses such as real-time data analysis. NPUs are also named among the purpose-built chips for AI. Beyond the processors themselves, AI hardware needs scalable cores, memory arrangements tuned for data locality, and strong interconnects. [4, 5]

Research directions in advanced semiconductors

The Nature collection lists several areas of current interest. In process technology these include nanosheet, forksheet and CFET nodes, as well as co-optimization of design, technology and systems. In packaging they include chiplet-based design, 2.5D and 3D integration, and wafer-scale computing. Another area is in-memory and near-memory computing, which uses emerging memory devices such as ReRAM, MRAM, PCRAM, FeRAM and FeFET. The collection also covers unconventional paradigms, namely integrated photonic, quantum and superconducting computing. A further topic is hardware-software co-design, for example model compression and approximate computing. [1]

Data centers, the edge and the market

The Semiconductor Industry Association reports that a single AI server rack holds over 4,500 packaged chips, made up of roughly 20,000 individual dies. It says chips make up more than 95% of such a rack's content value. Not only AI accelerators are needed. Data centers also depend on memory, networking and power-related chips. AI is also moving from central data centers to devices such as smartphones, smart cameras and industrial sensors. This calls for chips with high performance, low latency and good power efficiency. One industry article expected AI semiconductor revenue of $71 billion in 2024, a 33% rise over 2023. That figure is a forecast, not a measured result. [4, 7, 9]

A two-way relationship between AI and chips

A feedback loop is described. Chips provide the computing base for AI, and AI in turn is changing how new chips are designed and optimized. The Semiconductor Industry Association likewise expects stronger AI to help improve chip design and manufacturing. This outlook comes from industry sources, and the sources do not present independent measurements of it. [4, 9]

Frequently asked questions

Why can't ordinary processors simply be made faster for AI?

Performance gains from shrinking conventional processors no longer keep pace with AI's demands. Physical limits, heat management and rising costs add to the problem. [1, 5]

Which chip types are most often named for AI?

GPUs, TPUs, NPUs, FPGAs and ASICs are named. [4, 5]

Is there an upcoming event on this topic?

An event titled Semiconductors, AI & Future Computing, billed as a world conference, is announced for June 07–09, 2027 in Singapore. [2, 8]

Areas

Areas in Advanced Computing, Hardware & Semiconductors

Topics

Topics in Advanced Computing, Hardware & Semiconductors

  • AI-Driven EDA Layout Synthesis

    Artificial intelligence can take over and improve some steps of the electronic design automation (EDA) workflow, notably for integrated circuits and…

  • ASIC Verification Workflows

    Verification is the stage of chip development that takes the most effort, and there is disagreement on exactly how much.

  • Chip Floorplanning Optimization

    Floorplanning decides where the major building blocks of a chip sit before detailed placement and wiring begin.

  • Co-Design Approaches for AI Hardware

    Co-design treats algorithms, software and hardware as parts of one system that are developed together, instead of building each in isolation.

  • Compute Cluster Load Prediction

    Load prediction for compute clusters means estimating how much computing capacity upcoming workloads will need.

  • Compute Graph Partitioning

    Partitioning occurs in two neighboring senses: splitting graphs across processors, and dividing GPUs among workloads.

  • Cryogenic Quantum Hardware Monitoring

    Many quantum processors only work in extremely cold environments, so the refrigeration and wiring around them is a core part of the machine.

  • Embedded Firmware Security Analysis

    Firmware is the low-level software that sits between hardware and higher-level operating systems, so its security affects entire devices and networks.

  • FPGA Bitstream Optimization

    An FPGA is a chip whose internal logic can be set up again after manufacture, and the configuration bitstream is the file that defines that setup.

  • High-Bandwidth Memory Access Modeling

    High-bandwidth memory (HBM) sits right next to the processor in modern AI accelerators, and how a workload uses it often decides how fast the…

  • High-Speed SerDes Optimization

    High-speed SerDes blocks move data between chips by turning parallel data into a serial stream and rebuilding it at the receiver.

  • Low-Latency Inference Pipelines

    Low-latency inference means a model produces output quickly enough that a person or a downstream system notices no delay.

  • Multi-Chip Package Design

    Multi-chip package design places several chips or chiplets in one package, so that they work together like a single device.

  • Neural Network Hardware Mapping

    Hardware mapping is about how a neural network's computations and data are placed onto a specific chip.

  • Power Integrity Modeling

    Power integrity modeling studies how well a chip or board receives the power it needs, and how voltage noise and losses affect that supply.

  • Quantum Circuit Compilation

    A quantum program is a circuit, meaning a list of gates applied to qubits and followed by measurements.

  • Real-Time Signal Denoising

    Real-time denoising removes unwanted noise from a signal while the data is still arriving, so the compute budget is tight.

  • Thermomechanical Stress Simulation

    Thermomechanical simulation links heat transfer with structural mechanics.

Tools

Tools for Advanced Computing, Hardware & Semiconductors

All 22 tools in this industry

AI communities for Advanced Computing, Hardware & Semiconductors

Exchange, courses and practice around AI in this field — selected by us.

¹ Partner link: if you join through this link, we receive a commission; the price stays the same for you.