The G in GPU is for Graphics damnit

Original Article ↗ Hacker News Discussion ↗

Rendering, Lighting, and “Real” Graphics Workloads

  • Several comments note that basic raster graphics are conceptually simple until you add shadows, reflections, refractions, and global illumination.
  • Path tracing is praised as conceptually simple but computationally brutal; achieving low-noise, production-quality output requires many tricks, advanced sampling, and often PhD-level expertise.
  • Techniques mentioned include BVH acceleration structures, Metropolis light transport, and especially modern ML-based denoisers that combine multiple frames plus motion/depth data.
  • Even high-end ray-traced games still show artifacts (e.g., unstable reflections in motion), illustrating the gap between theory and real-time implementations.

Is NVIDIA Still a “Graphics” Company?

  • Debate centers on whether NVIDIA is now fundamentally an AI company, a “compute” company, or still a graphics company.
  • One side: revenue is now dominated by AI/datacenter, gaming is small, and hardware plus software (CUDA, AI libraries, ecosystem) are heavily optimized for AI workloads.
  • Other side: they sell hardware; calling them an AI company just because buyers run AI on it is like calling a knife maker a restaurant company.
  • Some argue past and present success reflect long-term investment in general-purpose parallel compute (CUDA, GPGPU, HPC) rather than luck or pure “graphics.”

Hardware Evolution Toward AI

  • Datacenter GPUs like H100/Blackwell are described as shedding traditional graphics features: no display outputs, limited raster hardware, focus on tensor/matrix throughput and low precision formats (FP4, etc.).
  • You can technically game on such parts, but performance is poor relative to consumer GPUs.

Market Structure: dGPUs vs APUs and Gaming

  • Many see high-end discrete gaming GPUs as a relatively small niche: most users are served by integrated GPUs/APUs in laptops, phones, and consoles.
  • Others counter that PC gaming is still large in absolute numbers; what’s shrinking is the fraction of players who chase ultimate FPS/visual fidelity.
  • Result: little room for new PC dGPU vendors; broader GPU competition lives in APUs (Qualcomm, Apple, Samsung, etc.).

CPUs vs GPUs and Programmability

  • Some discussion around claims that CPUs are “better” for graphics: consensus is that quality can be identical; GPUs win on speed.
  • The key difference is programmability and control: CPUs handle branchy, divergent code better; GPUs excel at massively parallel, regular workloads.

AI-Assisted Graphics

  • Multiple comments connect the idea of “AI doing the graphics” to existing features like DLSS (ML upscaling) and frame generation (ML interpolation).
  • Speculation goes further: future models might enhance low-detail scenes using higher-level understanding of geometry and materials, not just upscaling.