Hacker News Reader

Comment by bsenftner | Hacker News Reader

It is not so much a "chip that powerful" but a "chip that correct". Gaming GPUs are allowed errors because the rendered images are known to be visual output only, whereas an engineering GPU may be used for hard analysis, which cannot afford any rounding or other classes of the acceptable errors gaming GPUs are allowed.

View on HN

Do you have a source for those claims? It seems highly unlikely to me that certain types of cpu's or gpu's are allowed errors to cut costs. Rounding mechanics are complicated, but I'd say a bug is a bug. As mentioned in a comment by jackmott in this thread, "engineering purposes" also often actively reduce precision for increased performance.

Double precision floating point and the addition of ECC memory.

In the case of nVidia the quadro cards are the only ones sold as manufactured by the company and are incredibly high binned.

Lots of people on this thread are conflating a few things such as ML, but I work in an industry where we deal a lot with GPUs and its certainly a fact that flipped bits are significantly more common than many think. If you do any work that requires the output to be nonvolatile then you can run into issues.

Perhaps he is referring to ECC memory - high end pro model NVIDIA cards have it, and consumer cards don’t. Here is a paper showing that there are memory errors on the consumer non-ecc cards.

http://ieeexplore.ieee.org/document/5493404/?arnumber=549340...

Are you talking about double-precision floating point math? Or are you saying the gaming chips actually have issues?

It is just complete nonsense. In fact deep learning/AI chips used for neural networks usually have reduced floating point precision because they don't need it.

Deep learning is still a fairly new and small field, scientific computing workloads are usually mostly double-precision.

We don't need it because errors in any particular value aren't important - the calculations assume an amount of inherent error greater than that produced by the hardware.

nonsense

Reply on news.ycombinator.com