CPU vs GPU vs TPU vs NPU vs LPU: how 5 chip architectures trade flexibility for AI speed
Roger_M_Taylor · x · 2026-09-03
A visual explainer of the five hardware architectures powering AI today, each making different tradeoffs between flexibility, parallelism, and memory access.
- CPU: a few powerful cores for general-purpose logic and branching; poor fit for matrix multiplication
- GPU: thousands of small cores executing the same instruction on different data — parallelism maps directly to neural network math, dominating training
- TPU/NPU/LPU: further specialization with different tradeoffs
The quoted linked article gives LLM engineers the intuition behind how GPUs actually work, so techniques like quantization, speculative decoding, and continuous batching stop looking like arbitrary tricks.
More from Infra
- Nvidia acquires Hugging Face, the 'GitHub of AI,' for $13 billion — Hakan_Ozalp · 2026-09-03
- Figure to deploy 100,000 Vera Rubin GPUs in 2027 for home humanoid robots — Distinct-Question-16 · 2026-09-03
- Hidden China risks in America's multibillion-dollar AI data center boom are well known, insider says — pstAsiatech · 2026-09-03
- Nvidia nears $12.9B Hugging Face acquisition, about 86x revenue — sanjaykalra · 2026-09-03
- Reka and NVIDIA unveil real-time 30B video model: 720p at 24fps, 11.8x faster on one H100 — RekaAILabs · 2026-09-03
- AI Data Center 'Powered Shell' Stocks: Q2 Earnings Recap and What Could Reignite Interest — BenBajarin · 2026-09-03