GPU Programming

Why is double-precision (FP64) performance on many consumer gaming GPUs much lower than single-precision (FP32), for example one thirty-second?

  1. AGraphics and AI need at most single precision, so little die area is devoted to double-precision units
  2. BDouble-precision computation is physically impossible
  3. CDouble-precision computation is permitted only on CPUs
Show answer and explanation

Answer: A. Graphics and AI need at most single precision, so little die area is devoted to double-precision units

Double-precision units take several times the area and power of single-precision ones. Game graphics and AI inference are fine with single or half precision, so consumer GPUs include few FP64 units and spend the area on lower precisions. Scientific computing that needs double precision, such as weather or fluid simulation, uses data-center GPUs with strong FP64 performance.

  • A. Correct. Products for scientific computing strengthen FP64.
  • B. FP64 works, just slowly.
  • C. There is no such rule.
Practice this topic Back to topic

More GPU Programming questions

  1. What is a characteristic of the tensor cores (matrix units) in recent GPUs?
  2. In GPU mixed-precision computation, why are multiplications often done in a low precision such as FP16 while accumulation is done in FP32?
  3. Summing the same floating-point data on a GPU gives results whose last digits differ slightly from run to run. What is the most likely cause?
  4. What is a common guideline for choosing the number of threads per block in CUDA?
  5. What is the advantage of writing a kernel with a loop like the one above (a grid-stride loop)?