Why GPUs aren’t ‘graphics cards’ anymore

GPUs were built for games. Now they power AI, rendering, and cloud compute—and the name no longer fits

The term graphics card feels outdated. GPUs haven’t really been about graphics for years. The same chips that once rendered shadows and reflections now train large language models and run AI inference jobs. The name stuck, but the purpose didn’t.

The irony is that some of today’s most powerful GPUs can’t even display graphics. Nvidia’s H200 NVL, for example, has no display output—it’s designed purely for compute. Tests like those from LTT Labs make this obvious: when they compared the H200 to the RTX 5090, the gaming card excelled in raw speed but ran out of memory on larger models. The H200 kept going because of its enormous memory bandwidth. It’s not a “graphics” card. It’s an accelerator.

From pixels to parameters

Modern GPU work falls into two categories. Compute-bound tasks depend on core throughput and clock speed. Memory-bound tasks depend on VRAM capacity and bandwidth. Inference and training often land somewhere between the two. That’s why a 5090 can handle prompt processing quickly, but a model that pushes beyond its 32 GB VRAM will crash or slow to a crawl. The H200, with 141 GB of HBM3e memory, keeps performing under the same load.

This shift redefines performance. Speed still matters, but so does capacity. The GPU industry now splits between cards built for display and those built for data. It’s a quiet identity crisis playing out in hardware.

For developers exploring inference performance, LLM inference in production: a practical guide offers practical insight into throughput, latency, and batching.