Google colab worse than local instance

[WINDY_WINDWARD]

i was recently doing some trials on google colab and i found that its lot worse performing compared to my local server running an underclocked 1080ti (for power reasons)

here are screen shots for proof

https://ibb.co/njf3TbZ

https://ibb.co/vZ1CnJd

both instance are running the same datasets with the same settings

 

[chervonij]

@"WINDY_WINDWARD"

Dude, that’s logical.
K80 with one GPU core (as in a Colab) has 2496 CUDA cores. A 1080Ti - 3584 CUDA cores.
And if we take the overall performance, which is affected by the architecture of the cores, frequency of the cores/memory, etc., then the K80 has a performance of 4.113 TFlops, and 1080Ti - 11.34 TFlops.

In my opinion there is nothing to discuss.
There are three reasons to use a Colab.
First, if you do not want to pay for electricity.
Second, if you need your PC for other tasks, and not just make fakes.
And third, if you do not have a PC with a powerful GPU (like me)

 

[narangd]

@"WINDY_WINDWARD"

Dude, that’s logical.
K80 with one GPU core (as in a Colab) has 2496 CUDA cores. A 1080Ti - 3584 CUDA cores.
And if we take the overall performance, which is affected by the architecture of the cores, frequency of the cores/memory, etc., then the K80 has a performance of 4.113 TFlops, and 1080Ti - 11.34 TFlops.

In my opinion there is nothing to discuss.
There are three reasons to use a Colab.
First, if you do not want to pay for electricity.
Second, if you need your PC for other tasks, and not just make fakes.
And third, if you do not have a PC with a powerful GPU (like me)

hello. 
I think K80 has a performance of 8.71tflops and 4.29tflops on the Tesla K40.
I have rx580 (6.175tflops) and K80 is faster than rx580 while training.
so... maybe 8.71 is the right number? 

 

[chervonij]

@"narangd"
Hey.
See the specification carefully.
8TFlops is a K80 with 2 GPU cores and 24 GB of memory.
On the Colab you get access to only one GPU core from K80.

Your RX580 is slower by default because it runs on the PlaidML backend.
While K80 on pure Tensorflow.

 

[narangd]

@"narangd"
Hey.
See the specification carefully.
8TFlops is a K80 with 2 GPU cores and 24 GB of memory.
On the Colab you get access to only one GPU core from K80.

Your RX580 is slower by default because it runs on the PlaidML backend.
While K80 on pure Tensorflow.

oh I didn't know Colab get access to only one GPU core from K80
so even only one NVIDIA GPU(about 4tflops) core can beat my AMD GPU(6tflops)    
I can't wait to test my future upgrade NVIDIA GPU 2070 super!
thank you for the info and forgive my ignorance