Question GPGPU testing with normal graphic cards

[Greyguy1948]

Anyone testing some graphic cards like this:
GPU timing comparison
Xeon 12T/6C of Nehalem type is compared with Tesla 2070 GPU on matrix multiply. This is FP64 but FP32 should be more interesting.
No opt is used as 90 sec for a 3 GHz Xeon running 2000x2000 is really slow.

 

[Ken g6]

AnandTech does matrix multiply benchmarks. I think this is FP32:

 

[Greyguy1948]

Here we can compare Open CL and CUDA:
CompuBench

The new AMD 66A0:00 is better than Vega. Radeon VII maybe?

 

[Greyguy1948]

This is Tesla V100:
Tesla with HBM2

Totally out of budget for most of us but it could be intersting to see which benchmarks are depending of BW and which are not.

 

[Greyguy1948]

Some comparison of Titan GTX and Xeon E5-2665(8C/16T type Sandy Bridge EP):
GTX Titan and Xeon

 

[Greyguy1948]

Some tensorflow tests:
Tensorflow
Memory is important and 16 GB is much better than 8 GB

 

[Greyguy1948]

This is matrix multiply for some cards (FP32):
Matrix multiply

It is very much up and down like a switchback. I guess the cache in a GPU is not as sophisticated as in a CPU.