All geek questions in one place

Tesla double precision

I'm looking for information on how double precision is hardware implemented in tesla gpu. I read that two stream processors are working on the same double value, but I did not find the official paper from nvidia.

Thanks in advance. PPS Why are most GPUs calculated with a single precision (since colors can be saved as RR.GG.BB.TT, where each character is an 8-bit value)?

PS google it for me did not help

+3
source share
3 answers

Do not support double - this is not a storage format, as you said (RR.GG.BB.TT), but having its own built-in tools (and so dedicated equipment) for processing double operations (add, mul, madd, etc.) .

In any case, most GPUs support only one precision, because where most of the GPU market lies in the gaming market, gamers do not need double precision. Most gamers are also looking for good performance / prices. Implementing a DP is expensive in terms of a transistor (and TDP) budget, and if games do not use double precision, it makes no sense.

That's why you see ATI's high-end dual-GPUs (HD 59xx and HD 58xx, but not entry-level GPUs like HD 57xx or less).

@karlphillip: , , IEEE754 () , GTX 260, ATI NVIDIA IEEE 754-2008 .

, , IHV :)

+3

Tesla , . Tesla GPU Fermi , .

Fermi, . 9.

, .

+2

Wiki:

( , GTX 260 [12]) IEEE 754: - - , , NaN ; IEEE ( ), , ; / , .

, IEEE 754, , , .

+1

Source: https://habr.com/ru/post/1759147/

More articles:


All Articles