From the moment we turn on the computer and the Windows logo appears until we launch a game full of lights, shadows, and textures , everything goes through the same component: the graphics card . This component can be integrated into the processor or come in a dedicated format, and its mission is to transform data into images smoothly, accurately, and, in recent years, with a touch of artificial intelligence.
In this journey, we'll take a close look at the evolution of graphics cards, from VGA to GPU , the shift from early monochrome adapters to real-time ray tracing, and how all of this has impacted the Windows experience and gaming . We'll explore history, key technologies, APIs, manufacturers, buses, memory, buying tips, and even how to check your graphics card in Windows with dxdiag in just two clicks.
What is a graphics card and how does it work with the CPU?
A graphics card (or GPU, strictly speaking) is a processor specialized in floating-point operations , designed to execute thousands of parallel calculations that shape pixels. The integrated version is located within the CPU itself (iGPU/APU), while the dedicated version is installed on the motherboard via PCI Express and has its own VRAM , power supply, and cooling system.
The typical workflow on a Windows PC with games is as follows: the CPU prepares geometry (vertices), commands, and physics; the GPU organizes the scene (spatial ordering and clipping), and then executes the pixel/fragment shaders that provide color, materials, effects, and post-processing. Afterward, the signal is output via VGA, DVI, HDMI, USB-C, or DisplayPort to the monitor, which displays it at a specific refresh rate ( 50/60/120/144 Hz , etc.).
Dedicated graphics cards are typically much more powerful than integrated ones, which is why for editing, gaming , or AI, models with fast VRAM and high bandwidth are now prioritized. In gaming laptops, Max-Q chips optimize power consumption and temperature to deliver desktop-level performance with fewer watts.
From MDA and CGA to HGC, EGA, VGA and SVGA: the foundations
The starting point for PCs was IBM's adapters in the early 80s. The MDA (Monochrome Display Adapter) displayed only alphanumeric text (80×25) with 4 KB of memory and monochrome monitors. The controller read ASCII values, the character generator composed a raster matrix for each character, and the monitor reproduced it at around 50 Hz.
In 1981, the first color graphics arrived with the CGA (Color Graphics Adapter), which ushered in the RGB era on PCs: up to 16 colors (8 with two intensities) and resolutions such as 320×200 (4 colors) or 640×200 (2 colors). It wasn't perfect, but it put color on the map for home use.
Meanwhile, in 1982, the Hercules Graphics Card (HGC) appeared : monochrome, yes, but capable of 720×348 resolution and with 64 KB of memory. It allowed for crisp text (14×9 matrix) and a graphics mode that was remarkable for the time.
The next step was IBM's EGA (Enhanced Graphics Adapter), compatible with MDA and CGA, with 256 KB of memory and 16 colors at 640×350 chosen from a palette of 64. In addition, it made screen changes smoothly, reducing the annoying flicker typical of CGA.
In 1987, the industry embraced VGA (Video Graphics Array): 640×480 in graphics mode, 720×350 in text mode, 256 colors chosen from a palette of 262.144, and the key difference: an analog signal to the monitor. This is why VGA incorporated the famous RAMDAC (memory digital-to-analog converter), which operated at up to 450 MHz in later models. VGA also brought widespread resolutions of 800×600 and 1024×768, and refresh rates of 50/60/70 Hz.
From 2D to 3D: buses, chips and the first big revolution
The 90s brought two revolutions: buses and the leap to 3D. The VESA local bus standard gave way to PCI in 1993, with more compact cards from brands like Matrox, Creative, and 3dfx (Voodoo). Shortly after, AGP (x2, x4, x8) arrived to accelerate texture traffic, with theoretical peaks of up to 2,1 GB/s, a precursor to the current 16-lane PCI Express .
On the chip side, manufacturers like S3 (Trio, ViRGE), Rendition, Matrox, 3dfx, NEC PowerVR, ATI, and a then-nascent NVIDIA (RIVA TNT/TNT2) emerged. The first 3D APIs on PC were established with OpenGL (from Silicon Graphics), Glide (proprietary to 3dfx), and Direct3D within Microsoft DirectX , which on Windows would eventually dominate PC gaming.
The real turning point was the GeForce 256 (1999), the first chip with hardware T&L (Transform & Lighting), unifying polygonal 3D acceleration and offloading geometric calculations from the CPU. From then on, ATI named its family " Radeon ," and the modern rivalry was born.
Unified shaders, hot shaders, and the jump to DirectX 11
NVIDIA experimented with so-called "hot shaders ," running the shaders at higher MHz than the rest of the GPU (for example, 600 MHz for the GPU and 1.500 MHz for shaders on the 8800 GT). Meanwhile, 1 GB of VRAM became widespread , and bandwidth was significantly increased to avoid bottlenecks.
The GeForce GTX 400/500 and Radeon HD 5000/6000 series brought DirectX 11 , a higher shader count, increased bandwidth, and versions that, for the first time, reached 3 GB of VRAM (in NVIDIA's high-end range, albeit in a limited way). The rule was: add more compute units and memory to increase raw power.
GCN, asynchronous computing, and the end of hot shaders
AMD responded with GCN 1.0 (Radeon HD 7950/7970), an architecture ahead of its time that favored DirectX 12 and asynchronous computing. It also standardized on more generous VRAM (3 GB versus 2 GB in NVIDIA equivalents), a decision that would be noticeable with more demanding games.
NVIDIA, with Kepler (GTX 600/700), said goodbye to "hot shaders," tripled the number of shaders between generations (GTX 580 to GTX 680), and gained momentum in DX11, although its early DX12 support was lukewarm. Even so, with the GTX 780 Ti (2013), it doubled the number of shaders and managed to run games in 4K on Windows PCs with surprising ease.
The big leap came with Maxwell (GTX 900): much greater efficiency and performance per watt; a GTX 970 with 1.664 shaders outperformed the GTX 780 Ti with 2.880, in addition to having more VRAM (4 GB) and better compatibility with DX12 . It was one of the most beloved graphics cards for its balance and longevity.
With Pascal (GTX 10), NVIDIA fully embraced DX12 and Vulkan; a GTX 1070 (1.920 shaders, 8 GB) outperformed the GTX 980 Ti. The GTX 1080 Ti became a legend, with 4K performance still holding its own in classic rasterization.
Turing, RT Cores, and Tensor: Ray Tracing and AI in Windows and Gaming
The next major milestone came with Turing (RTX 20), which added two specialized blocks: RT cores (ray tracing) and tensor cores (AI and inference). From then on, a GPU was no longer just "shaders + textures + rasters" and could handle effects previously prohibitively expensive due to performance costs, integrating with DirectX Raytracing on Windows.
To offset the impact of ray tracing, NVIDIA launched DLSS , an AI-powered upscaling technology that failed in its first version but succeeded with DLSS 2 thanks to its temporal image reconstruction . AMD responded with FSR/FSR 2 , which doesn't use AI but offers good cross-platform performance.
Ampere and Ada (RTX 30/40), RDNA2 and RDNA3: efficiency, caches and chiplets
With Ampere (RTX 30), NVIDIA massively increased the number of shaders per SM; an RTX 3060 (3.584 shaders) had twice the number of cores as the 2060, improved RT/Tensor, and increased clock speeds. Then, Ada Lovelace (RTX 40) brought a huge leap in efficiency: a 110W RTX 4060 outperforms a 170W 3060 by around 20% in raster performance.
The RTX 4090 outperforms the 3090 by about 40% in raster graphics and is the only card that smoothly handles Cyberpunk 2077's Overdrive mode (path tracing). Furthermore, DLSS 3 introduces frame generation, alleviating CPU bottlenecks in Windows by interleaving synthesized frames on the GPU.
On the red side, RDNA2 doubled the shaders compared to RDNA, standardized 16 GB in the high-end range, boosted clock speeds, and added a large L3 cache block, the " Infinity Cache ," to reduce dependence on external bandwidth. It also introduced the first ray tracing units for DXR gaming in AMD.
With RDNA3 , AMD refined efficiency, added second-generation RT cores, incorporated AI accelerators, and, most importantly, introduced a novel multi-chiplet design : it maintained a monolithic die for the GPU and outsourced the L3 cache to chiplets. This reduces the silicon footprint, improves yield , and lowers costs.
Looking ahead, everything points to multi-chip module ( MCM ) GPUs with multiple interconnected GPUs. It's just a matter of time: the area of a single die no longer scales well in cost and complexity when we're talking about tens of thousands of shaders.
Essential components: GPU, VRAM, RAMDAC, VRM and cooling
The motherboard houses the GPU (calculating core with L1/L2 caches), VRAM ( textures, framebuffers, intermediate buffers), VRM (power phases with MOSFETs, chokes, and capacitors), and the cooling system . Dedicated graphics cards typically use 6+2 pin connectors because the PCIe slot only delivers up to 75W.
The historic RAMDAC converted digital data into an analog signal for VGA/CRT monitors. Although today everything is digital (HDMI/DP), understanding its role in the transition is key: its frequency determined image stability; advanced models reached 450 MHz.
In terms of heat dissipation, both turbine ( blower ) and axial flow (multiple fans pushing air over a finned radiator) designs are used. Blower fans are compact but noisy and less efficient; axial fans are the norm in custom models due to their superior thermal performance.
Video memories: from EDO/SGRAM/VRAM/WRAM to GDDR6 and HBM2
Before the modern era, graphics cards used EDO RAM and SDRAM , then SGRAM (graphics-optimized SDRAM), VRAM (dual-port read and write), and WRAM (faster than VRAM and with block acceleration features, ideal for windows in Windows ). This marked the transition from 300–800 Mbps to much higher bandwidths.
Today, GDDR6 and GDDR6X dominate : "DDR" memory with extremely high effective frequencies (14–21 Gbps) and buses from 128 to 384 bits, achieving massive bandwidths. AMD has used HBM2 (a bus of up to 2048 bits with 3D stacking), with lower MHz but a brutal bandwidth, useful in extreme bandwidth scenarios.
Classic memory-resolution relationship (2D era): with 512 KB, 1024×768 at 16 colors; with 1 MB, 1280×1024 at 16 colors or 1024×768 at 256 colors; with 2 MB, 1280×1024 at 256 colors and 1024×768 at 65.536 colors; with 4 MB, 16,7 million colors at 800×600 and higher resolutions became widespread. Today, for modern games, 4–8 GB is the reasonable minimum at 1080p–1440p.
Video ports: VGA, DVI, HDMI, DisplayPort and USB‑C
The analog VGA signal is outdated, but it's worth understanding DVI : DVI-D (digital only), DVI-A (analog only), and DVI-I (both digital and analog). HDMI 2.1 supports up to 4K@120Hz and 8K@60Hz; version 2.0 is limited to 4K@60Hz (8-bit). DisplayPort 1.4 allows 4K@120Hz and 8K@60Hz with DSC; DP is the preferred interface for high refresh rate monitors on PCs.
USB-C with Alt Mode DP/Thunderbolt 3 can output 4K@60 video and combine data and power. On modern Windows systems, it's common to see DP and HDMI coexist, and on laptops, USB-C is often used as a multi-purpose output.
3D APIs on Windows: OpenGL, Glide, DirectX and Vulkan
APIs are the "language" that the game speaks to the GPU. OpenGL (industrial and very capable) and Glide (a subset optimized for 3dfx) defined the 90s. Microsoft integrated the DirectX family (Direct3D) into Windows, which had a shaky start, but from DX8/DX11 onwards became the dominant standard on PC.
Additionally, there's Vulkan (Khronos), a low-level, cross-platform language rooted in AMD's Mantle. In practice, most major Windows games use DirectX 11/12 , with DXR for ray tracing; OpenGL/Vulkan coexist in specific engines and ports.
Buses: PCI, AGP and PCI Express
The bus defines the data path between the GPU and the system. PCI was the bridge of the 90s. AGP increased bandwidth and allowed access to system RAM (at the cost of latency). With PCIe x16 (3.0, 4.0, and 5.0), GPUs communicate directly with the CPU via 16 dedicated lanes. PCIe 3.0 x16 offers approximately 15,8 GB/s bidirectional bandwidth; 4.0 doubles that; 5.0 doubles it again. Today, in games, it rarely becomes saturated.
Performance and metrics: FPS, TFLOPS, TMUs/ROPs and overclocking
FPS determines smoothness: more FPS means a smoother feel, limited by the monitor's refresh rate (vertical sync via V-Sync/G-Sync/FreeSync). To see your GPU's "ceiling," disable sync and observe the frametime in benchmarking tools.
TFLOPS measure floating-point operations per second, a benchmark of raw power, but not the definitive one: architecture, caches, color compression, bandwidth , and drivers all play a significant role. TMUs (texture mapping/filtering) and ROPs (rasterization, blending, z-buffering, antialiasing) significantly impact the actual pixel and texture throughput.
GPU overclocking typically ranges from +100–150 MHz , and for GDDR6 VRAM, even +900–1000 MHz effective, with noticeable FPS gains if the GPU isn't bottlenecked by the CPU. Popular tools include MSI Afterburner, EVGA Precision X1, and AMD Adrenalin (WattMan).
Choosing the right graphics card, matching it with the CPU, and avoiding bottlenecks
For office tasks and multimedia, a modern iGPU (integrated Intel UHD/ARC or AMD Radeon Vega in APUs) is sufficient; investing in a dedicated graphics card isn't worthwhile. For 1080p gaming , a mid-range GPU with 6–8 GB of RAM and a 6-core CPU offer a great balance of price and performance.
At 1440p/4K, think high-end (more shaders, improved ray tracing, and ample VRAM). Remember: the CPU determines how much geometry/physics is fed to the GPU; lowering the resolution eases the load on the GPU, but barely reduces the CPU load. Settings that are CPU-intensive: object density, NPCs, simulation, physics; those that are GPU-intensive: resolution , textures, anti-aliasing, ambient occlusion, tessellation, and, of course, ray tracing.
Form factors and cooling matter: measure your case and choose between two- or three-fan designs, or even AIO liquid cooling in high-end models. Manufacturers (ASUS, MSI, Gigabyte, etc.) often offer factory overclocks and more robust VRMs .
Gaming and Max-Q Laptops
In laptops, the GPU is soldered and optimized (RTX/GTX Max-Q series ) with lower power consumption and slightly lower performance compared to desktops. They share GDDR6 VRAM, unified drivers for Windows, and support for technologies like DLSS and RT, prioritizing low temperatures and battery life, and offering options like the mux switch to improve performance.
How to find out which graphics card you have in Windows (dxdiag)
If you want to identify your graphics card in seconds from Windows, use the DirectX diagnostic tool :
- Click on Home.
- Open "Run" from the menu Home.
- Write dxDiag and click OK.
- When the utility opens, go to the tab Screen.
- Check under “Device” for the name of your GPU and the the memory available.
Manufacturers and ecosystem: NVIDIA, AMD, Intel and assemblers
Today, the market is led by NVIDIA (GeForce RTX) and AMD (Radeon RX). Intel has returned to the game with Arc in dedicated graphics cards and integrates graphics into most CPUs. In 2006, AMD acquired ATI; since then, it has combined CPUs and GPUs (APUs) and competes head-to-head in gaming and desktop PCs.
System builders (ASUS, MSI, Gigabyte, etc.) purchase GPUs and memory and design their own PCBs, VRMs, and heatsinks. Some add RGB lighting, dual BIOS, sensors, and overclocking profiles. In 2004, NVIDIA reintroduced SLI (multi-GPU) to boost performance, and cloud computing solutions like grid systems , which run graphics from remote servers, have existed, predating today's game streaming.
Looking at the entire journey, from the analog signal of a VGA to path tracing and AI-powered frame generation, it's clear that graphics cards have gone from being simple adapters to becoming massively parallel processors with a direct impact on Windows, engines, and games; the future reinforces specialization (RT/AI), chiplet designs, and efficiency, with the promise of higher fidelity and more FPS without skyrocketing power consumption.