Borrowed an H100 but couldn't draw a single frame — Why compute GPUs and rendering GPUs are different beasts
📝 Originally published (in Japanese) at forge.workstyle.tech . When someone says, "You need a GPU," they're actually referring to two distinct scenarios: Compute : Training and inference. As long as CUDA works, it's fine. Rendering : 3D rendering, gaming, video production. Requires OpenGL / Vulkan / EGL. Even within NVIDIA's GPU lineup, some models can only handle one of these tasks. I was…
When someone mentions needing a GPU, they are typically referring to two different situations: compute tasks like training and inference, which can run on CUDA-enabled GPUs, and rendering tasks like 3D rendering, gaming, and video production, which require OpenGL, Vulkan, or EGL support. NVIDIA's GPU lineup isn't a one-size-fits-all solution; some models are optimized for one purpose over the other.
During a project to create an AI avatar streaming system, the author encountered the "compute vs. rendering" limitation twice. First, using CPU rendering with SwiftShader proved insufficient, as rendering performance didn't improve despite lowering resolution. Second, allocating a single MIG slice from an H100 GPU didn't enable graphics APIs, despite the system having computational resources available. NVIDIA's MIG is designed for compute purposes only and doesn't support graphics APIs.
The author's solution involved renting a rendering-capable GPU from the cloud, specifically designed for graphics tasks. The RTX 4000 Ada and RTX 2000 Ada series GPUs are capable of handling 720p30 rendering at around 57-58 frames per second. The key takeaway is that compute and rendering GPUs serve different purposes; expensive GPUs aren't necessarily versatile, and a rendering-capable GPU is essential for tasks like 3D rendering and video production.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
