Do we have to identify tvm.nd.array context?

for example, if we use ctx=cuda,

both input datas are run correctly.

I saw that default tvm.nd.array context is cpu(0),

input = tvm.nd.array(shape)

but, I designate context “cuda”

input = tvm.nd.array(shape, ctx)

both inputs are run correctly, but it have small performance diffs.

Can we get better perfs with designated context like second one?

I want to know about principle of tvm data context