Hacker News Reader

I may be mistaken, it's my understanding that the video game cards are designed to be faster at the expense of accuracy, whereas the scientific computing cards are designed to have better accuracy at the expense of speed.

View on HN

Are the game cards not following IEEE floating point spec? I really should ask this on the nvidia cuda forums to get a really good answer.

Edit: Wikipedia says that the Tesla cards have 4X higher double precision performance than the game cards, so that could be the explanation.

That's one part of it, the other part is that subtle changes to the memory model will take a while to be taken advantage of, most code is now optimized for the 200 series chips (of which the 480 mentioned here has two).

Reply on news.ycombinator.com