what is compared in gpu demo

does 'paralleldemo_gpu_benchmark.m' compare a gpu to a single host core?

then the processing benchmark is biased since host processing with a lot of host cores should outperform gpu double precision by far

[Walter Roberson]

It depends on what hardware you have.

AMD Instinct MI300X Delivers 163.4 TFLOPS for FP64... and goes for about $US39000

NVIDIA H100 (SXM): 67 TFLOPS using its dedicated double-precision Tensor Cores... and goes for about $US31000

The double precision floating point performance of the newer GeForce GPUs is not especially high, as they are intended for gaming and AI, neither of which use much FP64.

Answer 1

[Walter Roberson]

No, that benchmark does not compare to host processing speed at all.

Answer 2

[Sean Sullivan]

The example ( https://www.mathworks.com/help/parallel-computing/measuring-gpu-performance.html ) is intended to allow you to measure some performance characteristics of your GPU and make some comparisons to your CPU, so I don't see how it could be biased. It makes some very general conclusions:

• Transfers from host memory to GPU memory and back are relatively slow.
• The GPU can read and write its memory much faster than the host CPU can read and write its memory.
• Given large enough data, a GPU can perform calculations faster than the host CPU.
• GPUs perform calculations faster in single precision than double precision, and often much faster.

If your CPU outperforms your GPU for your particular calculations, then use your CPU.