February 18, 2026
Math wars: nukes vs gamers
15 years of FP64 segmentation, and why the Blackwell Ultra breaks the pattern
Blackwell Ultra shakes up GPU math: fans cry foul, nukes get blamed
TLDR: NVIDIA kept strong double‑precision math off gaming GPUs for years, but the AI wave and Blackwell Ultra suggest that wall may crack. Commenters feud over whether this was market slicing, nuclear export rules, or silicon cost—while outrage over NVIDIA’s datacenter ban turns the debate into a wallet issue.
NVIDIA’s 15-year habit of making “science‑grade math” (double precision, aka FP64) crawl on gaming cards while single precision (FP32) sprints has the community in full popcorn mode. The article says AI flipped the script and Blackwell Ultra might blur the line between gamer gear and lab gear, and commenters are fighting like it’s a season finale. One camp cheers the history lesson: wtallis marvels that GPUs built for graphics accidentally birthed general computing, turning shader toys into supercomputers. Then jjmarr drops the spiciest bomb: FP64 isn’t just business—it’s “nuke-grade” and gets tangled in export rules like Adjusted Peak Performance, cue jokes about needing a background check to run a spreadsheet. The AMD crowd barges in with receipts: throwaway81523 points to the Radeon VII’s “oops, real FP64” moment and a GPU list that still embarrasses today’s gaming cards. Meanwhile, gdiamos says this isn’t a conspiracy—it’s silicon real estate: FP64 units are big, pricey, and gamers won’t pay for math they don’t use. The legal drama hits hard too: SubiculumCode rages over NVIDIA’s 2017 EULA banning GeForce in datacenters, calling it corporate lock-in with a lawyer badge. The vibe: half conspiracy thriller, half econ lecture, with memes about “release the double” and “RTX 5090: swole in gym math, weak in lab math.”
Key Points
- •RTX 5090 delivers 104.8 TFLOPS (FP32) and 1.64 TFLOPS (FP64), a 64:1 ratio not driven by technical limits.
- •Since 2010, Nvidia consumer GPUs’ FP64:FP32 ratio degraded from 1:8 (Fermi) to 1:24 (Kepler), 1:32 (2014), and 1:64 (Ampere).
- •From GTX 480 (2010) to RTX 5090 (2025), consumer FP64 rose ~9.65x, while FP32 rose ~77.63x.
- •Enterprise/datacenter GPUs maintained ~1:2 or 1:3 FP64 ratios historically and offered features like ECC memory and NVLink; price gaps grew from ~5x (2010) to >20x (2022).
- •Modern AI training favors FP32 and lower precisions; in 2017 Nvidia updated the GeForce EULA to bar datacenter use of consumer GPUs.