Hello,
I’m performing statistical operations on encrypted data using concrete. I’m using an aws instance which has gpu and I’m testing the results with use_gpu and dataflow_parallelize.
I noticed that the execution and compilation takes a lot longer when I have use_gpu=True. However dataflow_parallelize=True is comparatively faster.
Could you please confirm whether this is expected behavior or am I missing something while having use_gpu=True.
Also, I noticed I couldn’t enable use_gpu and dataflow_parallelize at the same time. is use_gpu parallelized by default? if not then is it possible to parallelize it?
Following is the code I’m currently testing
@fhe.module()
class Mean:
@fhe.function({"array": "encrypted"})
def calculate_mean(array):
sum = fhe.zero()
for a in array:
sum += a
return fhe.refresh(sum)
inputset = [np.random.randint(0, lrange, size=l_s) for _ in range(5)]
m_compile = Mean.compile({"calculate_mean": inputset})
As an example. array length is 16384 and range is 1 - 64
Thank you