How do graphics processing units work?
REFUSED: N/A — proceeding with note (sufficient grounded facts found: OECD Tier-2 source + The Hindu article as Tier-4 primary source).
Wait — disregard above line; proceeding directly with the study note.
How Do Graphics Processing Units (GPUs) Work?
1. At a Glance
- A GPU (Graphics Processing Unit) is a processor built for massively parallel, simple arithmetic (millions of small calculations simultaneously), unlike a CPU, which performs a smaller number of complex tasks sequentially/quickly [S1].
- GPUs have evolved from niche gaming/graphics hardware into core infrastructure of the digital economy, powering AI training/inference, scientific computing, and cloud data centres [S1].
- Relevant to UPSC because GPUs sit at the intersection of GS-III (Science & Tech, Economy) and GS-II (International Relations/regulation) — critical for India's semiconductor mission, AI strategy, and global tech-competition debates [S1][S2].
- Regulatory scrutiny of GPU-dominant firms (e.g., Nvidia) by the EU, France, and China signals GPUs are now a strategic/antitrust-sensitive resource, comparable to oil or rare earths in past geopolitics [S2].
2. Why in the News
- The Hindu (International print edition, 20 February 2026, Page 10) ran an explainer "How do graphics processing units work?" by Vasudevan Mukunth, prompted by GPUs' central role in the AI boom and mounting regulatory attention on chipmakers [S1].
- The OECD's Competition Committee (2024–2025 working documents) flagged GPUs, TPUs, and AI accelerators as chokepoints in AI infrastructure, with the European Commission opening preliminary probes into GPU/cloud markets and the French Competition Authority conducting dawn raids in a GPU-sector investigation [S2].
- China's competition authority has also opened a probe into Nvidia for potential breaches of its Anti-Monopoly Law; the EU separately reviewed the Nvidia–OpenAI transaction, citing high entry barriers in discrete datacentre GPUs [S2].
3. Background & Evolution
- 1999: California-based Nvidia Corp. marketed the GeForce 256 as "the world's first GPU," designed to improve videogame performance and visuals [S1].
- Over the following 2.5 decades, GPUs expanded from gaming/visual-effects use into general-purpose computing (GPGPU), scientific simulation, and — most significantly since the 2010s — AI/deep-learning workloads [S1].
- Key architectural milestone: NVIDIA's "Volta" architecture introduced Tensor Cores, dedicated hardware units for fast dense matrix multiplication, extending the CUDA parallel-programming model [S3].
- Related/competing designs: Tensor Processing Units (TPUs), which (unlike GPUs and CPUs, which follow the Von Neumann architecture) use a systolic array design optimised for matrix operations [S3].
4. Core Static Facts
| Item | Detail |
|---|---|
| Full form | Graphics Processing Unit [S1] |
| First GPU (marketed as such) | Nvidia GeForce 256, 1999 [S1] |
| Core design principle | SIMD (Single Instruction, Multiple Data) — many simple calculations in parallel [S3] |
| Contrast with CPU | CPU = few complex tasks, fast (latency-optimised, large caches, branch prediction); GPU = many simple tasks, parallel (throughput-optimised) [S1][S3] |
| Key GPU sub-unit | Streaming Multiprocessors (SMs), each with scalar cores + matrix-oriented cores (e.g., Tensor Cores) [S3] |
| Competing chip architecture | TPU (Tensor Processing Unit) — uses systolic arrays instead of Von Neumann architecture [S3] |
| Core mathematical operation | Matrix and tensor multiplication (e.g., c₁₂ = a₁₁b₁₂ + a₁₂b₂₂) [S1] |
| Regulatory bodies scrutinising GPU market | European Commission, French Competition Authority, China's antitrust regulator [S2] |
| International policy body flagging GPU bottlenecks | OECD Competition Committee (2024–2025 working papers) [S2] |
5. Multi-Dimensional Analysis
Economic - GPUs are now a bottleneck resource for the global AI economy — cost and scarcity of specialised chips (GPUs, ASICs) constrain AI infrastructure build-out [S2]. - High capital and software-stack barriers to entry in discrete datacentre GPUs entrench incumbents like Nvidia, raising pricing-power concerns flagged in merger review [S2].
Geopolitical/Strategic - GPU/AI-accelerator supply chains have become a strategic chokepoint, prompting export-control and industrial-policy responses among major economies (implicit in OECD's 2024–25 AI-infrastructure competition review) [S2]. - Simultaneous investigations by EU, France, and China into the same firm (Nvidia) show GPUs are treated as a matter of national economic security, not just commercial competition [S2].
Scientific/Technological - GPUs' SIMD/parallel design makes them uniquely suited to deep learning, where training involves billions of repetitive matrix/tensor operations [S1][S3]. - Specialised hardware evolution (Tensor Cores, TPUs' systolic arrays) shows a broader trend of domain-specific architectures replacing general-purpose computing for AI workloads [S3].
Ethical/Governance - Antitrust probes (EU preliminary investigation, French dawn raids, Chinese Anti-Monopoly Law case) reflect governance concern over market concentration in a technology now foundational to AI, defence, and economic competitiveness [S2].
6. Recent Developments (last 12–18 months)
- 2024–2025: OECD Competition Committee working papers (DAF/COMP/WD series) repeatedly analyse GPU/TPU/AI-accelerator market concentration and propose potential competition-policy responses [S2].
- European Commission: Opened preliminary investigations into markets for cloud services and specialised AI chips (GPUs, TPUs, accelerators) [S2].
- France: Competition Authority conducted dawn raids as part of a GPU-sector probe, following its Generative AI market study [S2].
- China: Opened a formal antitrust investigation into Nvidia under its Anti-Monopoly Law [S2].
- EU merger review: Scrutinised the Nvidia–OpenAI transaction, citing high barriers to entry in discrete datacentre GPUs (need for a mature software stack) [S2].
- 20 February 2026: The Hindu published an explainer situating GPUs' architecture and current controversies for a general/aspirant readership [S1].
7. Prelims Hooks
- The world's first GPU (as marketed) was Nvidia's GeForce 256, launched in 1999 [S1].
- GPUs use a SIMD (Single Instruction, Multiple Data) architecture, enabling massive parallelism [S3].
- CPUs prioritise latency (fast completion of few complex tasks); GPUs prioritise throughput (many simple tasks at once) [S1][S3].
- GPUs are composed of Streaming Multiprocessors (SMs), containing scalar cores and specialised matrix cores [S3].
- Tensor Cores were introduced in Nvidia's Volta architecture for fast matrix multiplication [S3].
- TPUs (Tensor Processing Units) differ from CPUs/GPUs by using a systolic array design rather than the Von Neumann architecture [S3].
- The core computational task in AI training is matrix/tensor multiplication [S1].
- The European Commission opened a preliminary probe into GPU and cloud markets as part of AI-infrastructure competition concerns [S2].
- France's Competition Authority carried out dawn raids in a GPU-sector antitrust investigation [S2].
- China's antitrust regulator opened a probe into Nvidia under its Anti-Monopoly Law [S2].
- The EU reviewed the Nvidia–OpenAI deal, citing high entry barriers in discrete datacentre GPUs [S2].
- Nvidia is headquartered in California, United States [S1].
- GPUs moved from being used mainly in videogames to becoming core AI/digital-economy infrastructure over roughly 25 years [S1].
8. Mains Relevance
- GS-III: Science and Technology — "Developments and their applications and effects in everyday life," Awareness in IT, computers, robotics; also Economy — infrastructure, industrial policy on semiconductors/AI hardware.
- GS-II: International Relations — global technology governance, antitrust cooperation among EU/France/China on strategic tech resources.
- Possible Mains question stems: 1. "Discuss how the architecture of Graphics Processing Units (GPUs) differs from Central Processing Units (CPUs), and explain why this makes GPUs central to the current artificial intelligence revolution." (GS-III) 2. "GPUs have moved from being a gaming peripheral to a strategic economic resource. Analyse the geopolitical and competition-policy implications of this shift, with reference to recent antitrust actions against major chipmakers." (GS-II/GS-III) 3. "Critically examine the challenges India faces in building domestic semiconductor and AI-accelerator (GPU/TPU) manufacturing capacity, and suggest a policy roadmap." (GS-III)
9. Related Topics to Study Next
- India Semiconductor Mission (ISM) / MeitY — India's domestic chip-manufacturing push, directly relevant to GPU/AI-chip self-reliance.
- National AI Strategy / IndiaAI Mission — India's AI compute infrastructure policy, which depends on GPU access.
- Export controls on advanced chips (US CHIPS Act context) — geopolitics of GPU supply chains.
- Competition Commission of India (CCI) and Big Tech antitrust cases — domestic parallel to EU/France/China probes on Nvidia.
- Cloud computing and data centre policy in India — GPUs are the backbone of cloud/AI data centres.
- Moore's Law and semiconductor fabrication (foundries, TSMC, node sizes) — underlying hardware trend context.
- Artificial Intelligence ethics and governance frameworks (NITI Aayog's AI strategy) — downstream policy implications of GPU-driven AI capability.
10. Common Errors / Trap Areas
- Confusing GPU with TPU: GPUs are general-purpose parallel processors (Von Neumann-based, SIMD); TPUs are Google's custom AI chips using a systolic array, not Von Neumann architecture [S3].
- Assuming GPUs were originally designed for AI — they were first marketed (1999) for videogame graphics, with AI use emerging only later [S1].
- Mixing up CPU vs GPU strengths: CPUs are NOT "slower" GPUs — they are optimised for different workloads (complex sequential tasks vs. massive simple parallel tasks) [S1][S3].
- Assuming only the US regulates Nvidia/GPUs — EU, France, and China have all opened independent investigations; this is a multi-jurisdictional issue, not a single-country one [S2].
- Attributing the GPU antitrust scrutiny solely to "monopoly pricing" — the actual concern cited is barriers to entry (capital + software-stack dependency) in the datacentre GPU market, per EU's Nvidia–OpenAI merger review [S2].
11. Sources
- [S1] "How do graphics processing units work?", Vasudevan Mukunth, The Hindu (International Print Edition, 20 Feb 2026, p.10) — https://www.thehindu.com/todays-paper/2026-02-20/th_international/articleGVVFK550M-13584719.ece — (tier: 4)
- [S2] "Competition in Artificial Intelligence Infrastructure" and related OECD Competition Committee working documents (DAF/COMP/WD 2023-2025 series) — https://www.oecd.org/en/publications/competition-in-artificial-intelligence-infrastructure_623d1874-en/full-report/component-7.html — (tier: 2)
- [S3] Technical/academic background on GPU–CPU architecture, Tensor Cores, and systolic-array TPU design, drawn from arXiv survey literature (e.g., "A Computational Model for Tensor Core Units," "Hardware Accelerators for Artificial Intelligence") — https://arxiv.org/pdf/1908.06649 — (tier: 3, supplementary technical reference)