I'm here with Kari Ann Briski, Senior Director for Accelerated Computing
and Software Product at NVIDIA. Kari, how are you doing today?
I'm doing good, thanks for having me.
You're welcome. I understand you're doing a keynote tomorrow - can you tell me
about what you're going to be talking about and what you want to drive home to the audience?
Sure, so tomorrow I'm sharing some
experiences we've had for end-to-end deep learning AI applications and deploying them in the cloud.
Last year, I actually came and talked when we had just launched the Volta V100.
At that time, when I asked people to put their hand up if they'd heard about, not many people did
Hopefully, this year is a big difference because we're in every major cloud provider, every computer maker ships Voltas.
Last year I also talked about training and inference libraries and architecture specific libraries for our chips.
But here, I'm gonna uplevel the picture and share some
experience we've had in taking AI neural networks for research
deploying them into the cloud and just kind of a little bit of a DevOps approach.
What perspective do you think you've gained from working with AI?
Well, actually just being humbled by the amount of great people that I work with
and actually the rate and the pace at which technology is advancing.
I was actually checking the numbers and when I first started at NVIDIA
we had just launched our first inference chip, the P4, and our first Tensor RT, which is our
runtime library for inference and since then and just about 18 months we've
improved performance 10X. We were only inferencing probably around 600 frames per second
frames per second and now we're about like 8,000 frames per sec
so it's pretty pretty exciting.
You're involved heavily with the technical side of the industry
what do you think are the greatest obstacles at the minute in terms of
the culture and just getting organisations to grips with AI and the tech?
Yeah, I hate to say there are a lot of obstacles
because you know AI still young but it's also mature I kind of like the analogy I
think it's a 16 year old kid about to drive you know it's like you can put it
on the road literally you can put it on the road but you know just the maturity
it since it's like going forward so fast and at such a rapid pace
having the tools to be able to keep up having your continuous integration
continuous delivery be able to stabilize has actually kind of been hard I think
for enterprises to be able to adopt AI in a stable way yeah just the amount we
call the Cambrian explosion of of neural networks so you know AlexNet was
only five years ago and that's considered a toy now you know it's a
shallow network that no one is really deploying in a real world application we
know that the deeper networks are the more accurate but they also require
more compute that's why the sort of tools are now coming up around
these like being able to prune networks how do I deploy these networks so just
kind of this ecosystem around AI is just starting to pick up to be able to
support the deploying of real neural networks into the real world
So where do you see that going in the next year or so?
I'll say I definitely see more tooling coming out right? Just being able to debug your neural network be able to see
where you have performance bottlenecks. Are you I/O bound or are you compute bound?
Being able to do your end-to-end latency maybe you only you have an application
already that you can only have final latency budget and maybe you already
have some sort of statistical model in there that you're going to replace with AI and
so you can only fit it into this tiny little latency budget so you need
compute, you know, accelerated compute to be able to make that neural network perform within that budget
So The AI Summit is here for the next two days, you were here last year,
what are you really hoping to engage with this time around?
Everyone! I just take it all in and learn from people's use cases, their pain points, and how we can
make life better how we can make it easier to use and adopt AI really just
from the bottom of the stack all the way to the top.
What kind of advice would you give to people - what should be their priorities when they're just starting out?
Just learn. You know, we always try and point people to our Deep
Learning Institute which is a really great start I mean I you know I took
some of the Coursera courses when I first started and just being it yes oh
yeah of course just being able to grasp and understand
take as many classes read and just dive in you know don't kind of sit and wait
you just my advice is just kind of start diving roll up your sleeves and get your
hands dirty right away so that you can learn what's right for you and your
enterprise that's it.
well Kari thanks for talking to me today.