The AI Revolution: Energy, Ethics, and Advanced Packaging

Send us a text

What happens when the world's thirst for artificial intelligence collides with physical limitations? The answer lies at the intersection of silicon innovation and packaging technology.

From the conference hall of ECTC 2025 in Dallas, we're joined by Sam Naffziger of AMD who reveals how AI has transformed from a peripheral topic to the dominant force in computing within just five years. The economic motivation is clear: creating machines that generate intelligence offers "uncalculable value" to humanity. Yet as Nafziger explains, today's large language models face significant limitations—they excel at pattern recognition but struggle with true reasoning, suffer from hallucinations, and require human verification for mission-critical applications.

Perhaps most surprising is Nafziger's revelation about "the data wall"—general-purpose AI models have essentially exhausted the high-quality training data available on the internet. This constraint is pushing AI development toward more sophisticated approaches involving reinforcement learning and models that check other models, gradually shifting from simple pattern recognition toward deliberative thinking that more closely resembles human reasoning.

The conversation tackles the looming concern of AI's energy consumption, projected to reach 10% of global power by 2030. Rather than viewing this as an insurmountable problem, Nafziger offers a compelling perspective: AI's ability to optimize countless processes—from transportation routing to crop yields and manufacturing efficiency—could ultimately lead to net energy savings despite its own substantial power requirements.

For the advanced packaging community, the message is clear: your work sits at the foundation of AI's future. As computational demands increase exponentially, innovations in thermal management, power delivery, and component integration will directly determine how quickly and effectively AI can evolve. 

Connect with Sam Naffziger on LinkedIn or visit AMD.com to learn more about their competitive AI solutions and join the conversation about technology that's reshaping our world.

Support the show

Become a sustaining member!

Like what you hear? Follow us on LinkedIn and Twitter

Interested in reaching a qualified audience of microelectronics industry decision-makers? Invest in host-read advertisements, and promote your company in upcoming episodes. Contact Françoise von Trapp to learn more.

Interested in becoming a sponsor of the 3D InCites Podcast? Check out our 2024 Media Kit. Learn more about the 3D InCites Community and how you can become more involved.

Episode Transcript

Françoise von Trapp (00:00):
This episode of the 3D Insights
podcast is sponsored by the IEEEElectronic Component Technology
Conference, organized by theIEEE Electronics Packaging
Society, ectc brings togetherthe best in packaging components
and microelectronics systems,science, technology and
education in an environment ofcooperation and technical

(00:20):
exchange.
Learn more at Ectcnet.
Hi there, I'm Francoise vonTrapp, and this is the 3D
Insights Podcast.
Hi everyone, this week we arerecording live from ECTC 2025 at

(00:50):
the Texas Gaylord Resort inDallas, texas, for the 75th
edition.
We're hearing a lot about themost advanced of the advanced
packaging technologies.
What's driving this right nowis power-hungry AI, still top of
mind for every engineer in thisindustry, and the keynote
speaker today, sam Nafziger, ishere to talk about emerging
trends aimed at addressing thedemand for high-performance

(01:12):
computing.
Welcome to the podcast, sam,thank you.

Sam Naffziger (01:15):
Excited to be here, Françoise.

Françoise von Trapp (01:17):
Okay, so before we dive in, you're with
AMD.
Can you just share a little bitabout your background and your
role there?

Sam Naffziger (01:23):
Yeah.
So I started out inmicroprocessor design with a
focus on power efficiencyimprovements and how to extract
the most performance per wattout of these devices and I've
started to lead thatcross-company from a power
efficiency perspective andsimultaneously have driven a lot
of our chiplet architectureapproaches and design, which is

(01:47):
the way to extract moreperformance out of the silicon,
the advanced silicon technologyprocesses, and of course that
involves deep package technologyengagements and advanced
development.
So I've ended up leading thearchitecture cross
productproduct for the companyand sponsoring long-lead
technology development.

Françoise von Trapp (02:07):
Now, AMD is pretty well known for its
high-performance computingprocessors, really targeted a
lot towards, I think, gaming.

Sam Naffziger (02:17):
We have a broad spectrum of products, right,
yeah, and gaming is where wefirst deployed our advanced
hybrid bond 3D in the CPU sideactually.
So gaming's been kind of breadand butter and it's a really
great market and veryenthusiastic customers.
Right but make most of ourmoney actually in the data

(02:39):
center.

Françoise von Trapp (02:39):
Right, ok, and that is a big deal right now
, as we see this explosion in AImodels.
You were talking this morningabout how fast AI has grown
since I think you were talkingabout since COVID.

Sam Naffziger (02:57):
Yeah, I mean, it was hardly a topic five years
ago when I spoke at ECTC, andyet now it's the topic, and the
reason is the economicmotivation of AI.
I mean, we actually are gettingmodels that can replicate many
aspects of human intelligenceand, of course, if you consider

(03:19):
the most valuable commodity inthe world, you know how did we
get all the comforts of modernexistence and cars and computers
and medicine?
Right, it's human intelligenceinventing stuff.
So if we can now inventmachines that can produce
intelligence, it's ofuncalculable value to the world,
and so that's why there's somuch excitement and hype about

(03:41):
it.
Now, the kinds of intelligencethat we're manufacturing are
imperfect.
Right, we're constantlyimproving, and that's what, you
know, makes us feel exciting isthe evolution, the pace of
development.
It far exceeds anything I'veever seen in the industry.

Françoise von Trapp (03:56):
Why do you think it had such a drastic or
rapid escalation?
I mean, it seems like once theydeployed the first versions of
chat GPT that's right.
It really started to take off,even though there's other market
spaces that aren't reallyconsumer facing.

Sam Naffziger (04:15):
Well, it, and it goes way beyond consumers.
The applications of AI and inscience and medicine and
robotics are going to be ofimmense economic value and
that's what's really driving it.
And, in fact, when you can haveI mean ChatGPT was such an
explosion in awareness becausethe capabilities just blew

(04:39):
people away.
Its ability to compose entiresophisticated essays and
synthesize books, provide adistillation and a summary of
complex technical treatises intoeasily consumable paragraphs
that would have taken days foran expert in the field to
synthesize down.
The models are amazing.

Françoise von Trapp (05:01):
I hesitate with tools like ChatGPT, though,
because it's only as good asthe data that it's training on
right, and the data has to beextremely accurate and on point,
and I think I feel like there'sa lot of people out there,
especially younger people, whoare, you know, using it to write
their papers, for instance.
I mean, I heard that chat GPTstarted out really smart and

(05:25):
that it's getting dumber.

Sam Naffziger (05:29):
Yeah, there is a corruption factor that canages
the erroneous or fabricated ones, because you know, as you're,

(05:56):
I'm sure you're aware you knowhallucinations are an issue with
AI.
It'll make up answers if itdoesn't actually know and
present them.
Yeah, communicate them as ifthey are authoritative.
So Every AI response needs tobe checked right, and putting it
in mission-criticalapplications to make decisions
is not a good idea at this point.

(06:17):
Right, because the models arenot reliable.
They're very impressive, but wecan't depend on the results.

Françoise von Trapp (06:26):
So you mentioned just now medical,
industrial, robotics as three.

Sam Naffziger (06:32):
Those are some of the top ones.
Right, okay, and agriculture,robotics as those are.
Those are some of the some ofthe top ones.
And agriculture I mean, ifyou'd say ai for science, it
encompasses a vast field of drugdiscovery and genomics analysis
and agricultural improvementswhich are extremely compelling.
The ability of ai to synthesizevast amounts of data and come

(06:54):
up with useful conclusions fromthat vastly more data than
humans can possibly absorb.
You know climate data andhistories of crop yields for
specific variants in certainregions and the soil types and
fertilizers, and you know justtalking about the agricultural
aspect and come up with a planfor crop rotation and the
appropriate farming techniquesthat will maximize yields and

(07:16):
minimize losses.
So just an example.
I mean, you know there's humanexperts that can do that, but an
AI can essentially, for thesespecific fields, become
superhuman in its ability toprovide those kinds of
recommendations.

Françoise von Trapp (07:30):
Okay, so one of the things you were
talking about in your keynotewas about running out of data to
train models.
Can you explain what you meantthere?

Sam Naffziger (07:41):
So for the large language models, you know the
general purpose.
Like ChatGPT you mentioned,they have been trained on the
compendium of Internet datathat's out there and slurping in
all the books and all of theanalysis.
You know everything.
But the model developers try tofocus on the high quality data

(08:02):
that can make the model moreintelligent versus just a bunch
of random numbers.
And yeah, the internet's prettymuch been tapped out now for
these huge model trainingexercises.
That is somewhat independent ofthe specialty fields, like I
just mentioned say inagricultural medicine?

Françoise von Trapp (08:20):
Is that because it's more enclosed the
data that you're feeding?
This AI engine is alreadyqualified.
You know, it's not justscraping random data off the
Internet, they're actuallyfeeding it.

Sam Naffziger (08:32):
High-quality data .

Françoise von Trapp (08:33):
It's like high enclosed, like encapsulated
data that's not been corruptedby any other input, right?

Sam Naffziger (08:41):
Right, right, model contamination absolutely
can degrade the intelligence ofthe AI.
So that's a very importantfield.
But yeah, we have to someextent you know, for the general
intelligence applications hit adata wall.
And actually it's a good thingbecause now we're leaning into

(09:05):
new approaches that leveragereinforcement learning,
approaches that have feedbackloops and models, checking
models, and we often put humansin that loop as well human
reinforcement learning toimprove the quality responses,
to grade the responses, which isthe better response out of the
set.
And now, when we can automatethat with multiple AIs checking

(09:28):
each other and generatingsynthetic data, we've been
making significant strides inthe ability of these models to
actually reason, not justregurgitate answers.
So the initial LLMs, the GPTsand Geminis they're good at
using that vast trove ofInternet data.
They're trained on to producethe most credible response to a

(09:51):
given query.
But it's essentially justpattern recognition.
It's not really thinking.
It's just like give it aresponse and boom, here's the
answer.
It's kind of like system onethinking in the brain where you
can recognize faces reallyquickly.
But if you start thinkingthrough, okay, if I, if I see
John and I last saw him herewhat's the right way to respond?

(10:13):
Make a connection with John.
You know that's, know that's areasoning thing.
A system two and we're onlyjust starting to get models that
can do that.
More deliberative thinking.

Françoise von Trapp (10:23):
So I know we're limited on time, so I just
wanted to ask you two things.
First of all, one of the thingsthat we're hearing that people
are concerned about is theamount of energy that AI is
consumed, and there'sprojections that by 2030, 10% of
the world's energy will go topowering AI.
So I guess maybe I'm in themiddle of an existential crisis

(10:47):
around this.
So I mean, it's too late.
You know the genie's out of thebottle, but should we be
rolling this out before we'vesolved the energy problem?

Sam Naffziger (10:58):
Yeah, that's a fine question and I think, yeah,
people have every right to beconcerned about the energy
consumption of AI, because itdoes appear like it will
outstrip supply and powerlimitations become a very real
cap on the amount of AI we candeliver.
But I guess I would turn thataround a bit.
If you think about what we areachieving with AI, we are

(11:21):
developing machines that cansolve the world's hardest
problems and actually invent newapproaches or identify, I'd say
, optimal approaches totransportation, routing, to
minimizing power consumptionacross a myriad of industries,
to providing better crop yields,reducing pollution, improving

(11:42):
gas mileage, countless things.
So actually I view AI as a.
It's kind of like.
You know it's a compoundinterest return.
Investments in AI are going toimprove human productivity,
quality of life, health andactually reduce energy
consumption in net.
Even if itself is consuming alot of power, the intelligence

(12:05):
we're generating is going to beharnessed for vastly more
productivity.

Françoise von Trapp (12:10):
So maybe it's limiting the frivolous use
of AI and focusing it on theareas where it really can make a
difference.

Sam Naffziger (12:18):
It'll reduce inefficiency, okay Right.

Françoise von Trapp (12:21):
So just last question for you what does
AMD need from the advancedpackaging community to solve
these challenges that you talkedabout in your talk, and I think
up there was thermal issues.
I think maybe, yeah, had energy.

Sam Naffziger (12:38):
Yeah, thermal, you know.
Just power efficiency in general, power delivery, getting the
heat out, getting power in thepackage community is absolutely
foundational to achieving thenext wave of growth in AI.
Like I said, the economic demand, the potential of AI to make
the world better in countlessways, make our manufacturing

(12:58):
processes more more efficient,as well as medical and the
health benefits, drug discovery,all these things.
So the more intelligence we cangenerate with AI, I believe is
better for humanity, even though, like any tool, it can be used
for good or ill.
I believe, by and large, we'lluse it for good and we try to

(13:21):
marginalize the corrupt uses.
And the ability to generatemore.
Ai is limited by power, and thepackage community has a huge
role in providingpower-efficient connectivity for
the silicon chips that are atthe core of those compute
activities, right?
So, whether it's the memory,whether it's the compute devices

(13:43):
, the GPUs, the accelerators orthe networking, getting those
components closer together withthe most energy efficient
connectivity possible, with thebest heat conductivity, the
lowest resistance for powerdelivery, all of those sorts of
problems are going to enable usto develop AI faster and more
effectively, which I believe isa net good.

Françoise von Trapp (14:05):
Well, thank you so much for your time.
I appreciate it.
Can we connect people with youon LinkedIn?

Sam Naffziger (14:11):
Oh, absolutely yes, I'm on LinkedIn, and AMD.
com has great research about ourcompany's particular AI
solutions, which are verycompetitive.

Françoise von Trapp (14:24):
Okay, great .
Thank you so much.
Next time on the 3D InCitespodcast.
We wrap up our coverage of ECTC2025, talking with 3D Insights
member companies about their keytakeaways from this year's
event, what they were showcasing, and also some of their
memories of ECTC's past and whatthey hope to see in the future.

(14:47):
There's lots more to come, sotune in next time to the 3D
InCites ast.
The 3D InCites Podcast is aproduction of 3D Insights LLC.