welcome to this session of The Energy & AI podcast (subnetwork of DSR sponsored by Japan's NEDO hosted Professor David Sandalow) supported by the embassy at www.uae-ebassy.org or search for UAE- US tech cooperation --1 -- we thank them for their support and we look forward to it developing and growing over time because the issue is so important
in this podcast we welcome jensen huang to the AI Energy and Climate Podcast thanks for
1:19
joining us and I'm looking forward to
1:24
discussing a range of topics including the potential benefits of AI power demand for AI your plans at Nvidia and
1:31
more and maybe let's start with the potential benefits of AI what do you see as some of the most important ways that
1:37
artificial intelligence can contribute to solving energy and climate change problems
1:43 Jensen Huang" well artificial intelligence has the ability to understand uh data understand
1:49
if you will information at a scale no humans can; and multimodal information from temperature to wind
2 speed to pressure etc all at the same time; and it could study information
2 across a very large scale very large longitudinal scale and so so the ability to understand information and and then be able to predict the future quite an extraordinary power
but let me take a step back and and say something about about how we got here - as you know we at nvidia are now a third of a century into inventing a new way of doing computing called accelerated computing and and we reduced the amount of energy required to do computation by 25X or more in order to increase performance; the fundamental way of increasing performance is to reduce computational energy ; together with this our partners have supported debel;oping new way of doing computation called machine (dee) learning and in turn we have
3:33
this new way of doing AI and of course a lot of the conversation we're going to have is that there's two phases of AI:
there's the learning phase we call it the training phase of the model
and then there's the inferencing
3:46
model which is the application phase
it's a little bit like humans you know we we have to go spend years pre-training going to school and then afterwards we get to go apply it and inference
4:01 both inferencing and part training requires a lot of energy but the good news is that
4:09
most of most of the world's computation breakthroughs for humanity really going to be about inference where compuation energy can be low so we'll be able to talk about that during this talk
david: that's interesting your comments about driving energy consumption down are
4:28
counter to most of the narrative that we hear these days all this attention about how AI is driving up energy demand so
4:35
power demand could you just say say a word about what you mean about how AI is driving energy consumption down
jensen yeah you
4:41
have to look at AI in its entire longitudinal aspect you know because the
4:47
goal of AI is not to train the model and these data centers that we use to train the model are quite large and you you uh
4:55
you you you apply basically three things uh uh you have to invent a model and
5:01
that model is is uh has to have the ability to uh learn the uh predictive
5:07
features from the data that you're presenting and so it could be a lot of videos or it could be um large amounts
5:14
of uh telemetry data that's taken you know about about weather or the climate um it could be it could be uh uh a whole
5:22
bunch of words that you want to learn something about the language and so you collect an enormous amount of data and
5:28
you have a you have a model that you have to train and it takes the computation takes a computer to uh
5:35
iteratively uh figure out how to predict uh uh understand the the features and
5:42
how to predict the outcome and so that takes a large computer to do but the goal of
5:48
course is not just to train the model the goal is to use the model now once
5:53
you train a large model and the model is quite large you know it's trillions
5:58
parameters large but then what you do is is you've now created if you will a professor model you take this professor
6:05
model and you distill it down into a whole bunch of little student models and so the actual model that you use is not
6:12
that big it could fit in a cell phone it could fit in a PC it could fit uh in a self-driving car for example just one
6:18
chip on a self-driving car um you know and and the response time is really good it's really fast just as you know when
6:25
we're using chat GBT or Gemini the response time is super fast and because the model has been it trained in a very
6:32
large model but it distilled it into smaller medium small tiny models now of
6:38
course these tiny models have the skills of the large model but not the general
6:43
general generalization capability um of the large models and so it tends
6:49
to be a little bit more brittle it tends to be more specific but it has the capabilities to you know to perform
6:54
whatever task that it was fine-tuned to do
and so in the end what you have is you you train a large language model but
7:01
you inference a whole bunch of smaller models and and those small models require a lot less energy let me give
7:07
you one example so we trained uh trained a large model to to u predict uh
7:13
weather and and um uh the models that it learned from are principled simulation
7:20
models and so these are you know classical um uh physical physically based um
7:27
principled physics simulations and we would use we would use those models and observable observed
7:35
data to uh train a large langu a large simulation model large physics model physics space model but when we're done
7:43
we predict weather 10,000 times less energy instead of using a
7:50
supercompuert to predict the weather and this these supercomputers are running 247 trying to predict tomorrow's weather
7:56
and next week's weather now we could use an AI model that predicts tomorrow's weather next week's weather 10,000 times
8:04
less energy
david: i think that's a point that's not widely understood that these new weather models actually use less
8:09
energy than some of the old physics based models do and and that's had real applications in solar and wind farms and other places ...
jensen: yes david that's a big idea - more at video above
===
uea major sponsor of ity plenipotentiary 4G to 5g development 2015-20 while itu led by houlin zhao, and while jim kim, jack m,a fei-fei li. melinda gates helped launch ai for good
====
the emirates minister fir ai, alolama briliant interview - one of first data sovereignty jensen huang
============
