How the RTX 5090 Actually Works (Visually Explained)

2026-02-11 · Source: Bug · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Semiconductor Manufacturing · Depth: Intermediate, medium

Summary

Nvidia's RTX 1590 GPU, featuring the GB202 die, integrates 92.2 billion transistors on a 750 mm² silicon wafer, representing a significant semiconductor engineering challenge. The chip utilizes the Blackwell architecture, organizing transistors into 170 Streaming Multiprocessors (SMs), each containing 128 CUDA cores, fourth-generation RT cores, and fifth-generation Tensor cores, totaling 21,760 cores. Manufactured by TSMC using extreme ultraviolet (EUV) photolithography, the process involves etching patterns onto high-purity silicon wafers. To mitigate manufacturing defects, Nvidia designs the GB202 with 192 SMs but configures the consumer RTX 5090 to use 170, disabling faulty units to maximize yield. The GPU is mounted on a modular, three-piece PCB with 32 GB of GDDR7 memory connected via a 512-bit bus, powered by a 30-phase voltage regulation module delivering 575W, and cooled by a 3D vapor chamber and dual axial flow-through fan system.

Key takeaway

For Computer Vision Engineers evaluating high-performance hardware, the RTX 1590's Blackwell architecture and advanced manufacturing demonstrate how defect tolerance and sophisticated cooling enable extreme transistor density. You should consider the implications of its 575W power draw and 32GB GDDR7 memory for your project's power budget and data throughput needs, especially for large-scale AI model training or real-time ray tracing applications.

Key insights

Advanced GPU manufacturing combines intricate architectural design with precise photolithography and defect tolerance.

Principles

• Modular design enhances signal integrity and cooling.
• Defect tolerance improves manufacturing yield.
• High-purity silicon is critical for chip functionality.

Method

Chips are created by carving patterns into silicon wafers using EUV photolithography, followed by testing and disabling defective components to ensure product reliability and maximize usable output.

In practice

• Implement modular PCB designs for thermal management.
• Design for defect tolerance in complex silicon.
• Utilize advanced cooling solutions for high-power GPUs.

Topics

• NVIDIA Blackwell Architecture
• High-Performance GPUs
• Semiconductor Manufacturing
• Tensor Cores
• Photolithography

Best for: Computer Vision Engineer, AI Engineer, Machine Learning Engineer, AI Architect

Related on AIssential

• Now Generally Available, NVIDIA RTX PRO 5000 72GB Blackwell GPU Expands Memory Options for Desktop Agentic AI — The NVIDIA RTX PRO 5000 72GB Blackwell GPU expands desktop AI capabilities with increased memory for larger models.
• This new 3D chip could break AI’s biggest bottleneck — Monolithic 3D chip integration, produced in a U.S. foundry, significantly boosts AI hardware performance by overcoming memory and miniaturization walls.
• Evaluating CUDA Tile for AI Workloads on Hopper and Blackwell GPUs — CuTile offers high productivity and peak performance on datacenter Blackwell, but lacks portability and consistent performance across other architectures.
• This NPU Is 5,000% Faster Than A GPU | Photonic Chips Explained — Photonic MPUs offer extreme energy efficiency for AI inference by using light waves, overcoming silicon's resistive heat limitations.
• Gemma 4 on Old GPUs: Why a $700 Used Card Beats $20,000 of Professional Hardware — Used RTX 3090s offer optimal price-performance for local LLM inference due to MoE architectures and high VRAM/bandwidth.
• H100 to B200: What Actually Changed — NVIDIA's Blackwell architecture, featuring the B200 GPU and GB200 NVL72 system, represents a fundamental shift from the H100, moving beyond individual GPU performance to datacenter-scale AI infrastructure.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Bug.