Seven Things You Might Not Know about Numba

Technical Blogs & Events Technical Blog
1

Originally published at: Seven Things You Might Not Know about Numba | NVIDIA Technical Blog

One of my favorite things is getting to talk to people about GPU computing and Python. The productivity and interactivity of Python combined with the high performance of GPUs is a killer combination for many problems in science and engineering. There are several approaches to accelerating Python with GPUs, but the one I am most…

2

Wow, this is super cool!!! Thank you for sharing!

3

Hi All,

Here is what I did with Numba:

I just got published an blog article "High Performance Big Data Analysis Using NumPy, Numba and Python Asynchronous Programming" in Dataconomy media (http://dataconomy.com/). Here is the link: http://dataconomy.com/2017/...

Let me know what you think?

Thanks

Ernest Bonat, Ph.D.
Senior Software Engineer
Senior Data Scientist

4

Did you run it on a GPU, or just the CPU?

5

Hi Mark,

I run it in CPU laptop! (32 RAMs)

Thanks

Ernest Bonat, Ph.D.
Senior Software Engineer
Senior Data Scientist

6

Did you consider compiling it to run on GPU?

7

Yes, I would like to do that. I wish I can have a GPU laptop for it!

Thanks

Ernest Bonat, Ph.D.
Senior Software Engineer
Senior Data Scientist

8

Hi Stanley,
Looks like a typo snuck into your Cuda C++ clamp example, patched thusly:

{code}
__host__ __device__ float clamp(float x, float xmin, float xmax) {
if (x < xmin){
return xmin;
- } else if (x > xmin) {
+ } else if (x > xmax) {
return xmax;
} else {
return x;
{code}

9

Thanks! Fixed.

10

Probably a very late response but:
Regarding “Numba + Jupyter”: I find that Jupyter somehow limits the sched_affinity such that Numba is only using 2 cores when 96 are available on my cluster node. I’d love to know how to get around that. Just setting “os.environ[‘NUMBA_NUM_THREADS’] = ‘96’” within the notebook doesn’t seem to do it; still only uses 2.