Explore GPU

Sanity Check #1:



In [ ]:

    
%%bash 

nvidia-smi



In [ ]:

    
%%bash 

xla_device_test &> xla_device_test.log

tail -3 xla_device_test.log



In [ ]:

    
%%bash 

cat /root/src/main/cuda/SumArrays.cu

### EXPECTED OUTPUT ###
...
*** Awesome!  The GPU summed the arrays!! ***
...

Note the execution time.



In [ ]:

    
%%bash 

sum_arrays

watch -n 1 nvidia-smi

Note: Don't go higher than 10!

Otherwise the following may happen:



In [ ]:

    
%%bash

# Don't go above 10!!
for _ in {1..10}
do
  sum_arrays > /dev/null 2>&1
done

echo "...Done!"

We lower overall execution time using async, stream-based memcpy



In [ ]:

    
%%bash 

cat /root/src/main/cuda/SumArraysAsyncMemcpy.cu

### EXPECTED OUTPUT ###
...
*** Awesome!  The GPU summed the arrays!! ***
...

Also, note the lower execution time due to async memcpy.



In [ ]:

    
%%bash 

sum_arrays_async_memcpy



In [ ]:

    
%%bash

# Don't go above 10!!
for _ in {1..10}
do
  sum_arrays_async_memcpy > /dev/null 2>&1
done

echo "...Done!"



In [ ]: