Floating-Point Hardware From Pixar to AI Chips
June 29, 2025

One arithmetic demand runs from cinematic rendering to modern artificial intelligence: perform enormous numbers of floating-point multiply-add operations and keep the data moving around them.

Brian Greenforest traces that demand from Ed Catmull’s graphics research and Pixar’s RenderMan through SGI-era workstations, programmable shaders, NVIDIA GPUs, and deep learning.

Movies Made Floating-Point Throughput Visible

Catmull’s work on z-buffering and texture mapping helped establish computational graphics. The Lucasfilm Graphics Group and later Pixar turned physically rich images into a production workload, while RenderMan gave artists a programmable description of surfaces and light.

DEC and SGI workstations, custom ASIC suppliers, and specialized geometry and raster hardware advanced around that demand. The history reflects an industry-wide convergence rather than one corporate handoff.

Shaders Carried the Arithmetic Into AI

Programmable GPU stages turned graphics hardware into a general platform for regular parallel math. Inigo Quilez helped spread fragment-shader thinking through Pixar work and ShaderToy, making spatial kernels available to a global creative community.

AlexNet’s 2012 GPU training demonstrated how well the same throughput architecture fit neural networks. Graphics and AI now share a hardware center of gravity: FMA arrays, memory bandwidth, programmable kernels, and relentless scale.

Trace Floating-Point Hardware From RenderMan to AI

The linked hardware history follows FMA-heavy rendering through programmable shaders and into modern AI accelerators.

The Shader Before the Programmable GPU: RenderMan, Graphics, and AI Hardware · Cheap Pixelless Textures With 2D SDFs · Why Open-Sourcing ASIC FMA Is Hard

Originally posted on LinkedIn

Brian Greenforest · (2025-06-29 03:06:47 UTC)

Open the original LinkedIn post · LinkedIn activity 7344924266917711872

LinkedIn status when archived: Visible to anyone on or off LinkedIn.

In the late 1970s, Ed Catmull at the University of Utah developed foundational graphics methods like z-buffering and texture mapping. By 1979, he joined George Lucas at Lucasfilm, forming the Graphics Group. Their work—eventually named RenderMan—required computers able to handle massive floating-point multiply-add (FMA) operations per frame. DEC and later SGI supplied the advanced workstations for this. LSI Logic, as ASIC supplier, manufactured custom chips for SGI and DEC, with future Nvidia founder Jensen Huang working at LSI, gaining direct insight into these compute needs. In 1986, Steve Jobs acquired Catmull’s group, renaming it Pixar. Pixar’s RenderMan software, released in 1988, let filmmakers program surface appearance and light. It became the gold standard for photorealistic animation and VFX, used in films like Toy Story, Jurassic Park, and The Lord of the Rings. RenderMan was and remains licensed to nearly every major studio, including ILM, DreamWorks, and Weta Digital. The RenderMan division at Pixar/Disney is estimated to generate tens of millions in annual licensing revenue, with pricing for full-feature licenses for major studios running from tens to hundreds of thousands of dollars per production. During the 1990s, SGI and Pixar coordinated to optimize hardware for RenderMan. LSI Logic continued supplying custom floating-point engines until Jensen Huang left to found Nvidia in 1993. By 2001, Nvidia shipped GeForce 3 with programmable shaders, applying the same FMA-heavy computation that powered Pixar’s rendering to the real-time world. Procedural graphics experts like Inigo Quilez brought fragment shader techniques to Pixar (notably for Brave), and later spread these ideas globally via ShaderToy. In 2012, Ilya Sutskever and colleagues trained deep neural nets on Nvidia GPUs, using the same floating-point architecture developed for graphics. Today, language models like GPT run on this hardware. In 2024, Catmull joined Odyssey, linking AI and graphics once more. The thread through all of this is the industry’s demand for scalable floating-point multiply-add operations, first for movies, then for AI. RenderMan, its clients, and the supporting hardware have shaped the evolution of both visual effects and artificial intelligence.

Original LinkedIn media

Media shown with Brian Greenforest's original LinkedIn post
Photo attached to the original LinkedIn post. Open the saved full-resolution image.

Comments added by Brian Greenforest on LinkedIn

These 2 comments were also preserved verbatim from Brian Greenforest’s LinkedIn data export or the public post page.

Comment 1 · (2025-06-29 03:22:10 UTC)

Image credit: Silicon Graphics geometry engine board featuring custom VLSI Technology ASICs and Xilinx FPGAs, used for hardware-accelerated floating-point operations in SGI workstations.

View the LinkedIn post

Comment 2 · (2025-07-11 15:27:20 UTC)

Chris Thornborrow, my ShaderToy friend Wyatt Flanders made it fun at the right time in 2019 for me personally, letting the whole story unfold beautifully, experimenting with cellular automata first. It was much later in 2021, when I finally learned how SDF works, and just recently finally found how amazing and cool Inigo was giving the world this beautiful social network of digital artists who render things the way Pixar does. Then came same rendering on FPGAs, one pixel at a time, for retro computing in Atari style. Right now I'm listening Walter Isaacson's Innovators (chapter 6) about how Atari became reality of hyper-simplified and fun user interfaces, BECAUSE PDP was so expensive. Look, ma: no CPU! 😍 https://www.reddit.com/r/proceduralgeneration/comments/esmaoj/interview_with_inigo_quilez_lover_of_equations/

View the LinkedIn post