00:00:00
AI is gradually making its way into every
industry and now it's reshaping computer
00:00:06
chip design Google's DeepMind announced the
major breakthrough in AI driven chip design
00:00:12
with its new AlphaChip according to the paper
AlphaChip is compressing month of work into hours
00:00:20
and eventually generating better chip designs
this sounds huge if true let's take a look for
00:00:27
nearly a decade I was deep into designing silicon
chips picture this billions of transistors placed
00:00:34
on a tiny piece of silicon all connected by 30
miles of wires literally solving this is like
00:00:42
building a very complex puzzle where every
single piece has to fit perfectly to make it
00:00:47
work and nowadays it's hard to imagine that
back in ' 70s this job was done by hand back
00:00:54
then circuits were literally drawn on a piece of
paper but as designs were getting more complex
00:01:00
featuring more transistors and more
interconnects chip makers started to
00:01:04
develop software tools to do their job as of
today no one is placing and interconnecting
00:01:10
billions of cells by hand anymore these days we
used so-called EDA Electronic Design Automation
00:01:17
Tools for that these tools are really great
at automating many aspects of the chip design
00:01:23
flow they run lots of math on the background
to find a way to place billions of transistors
00:01:29
and interconnect them in the most efficient
way and the top players in the EDA market
00:01:35
are Synopsys and Cadence I own their stock and
they play a critical role in the semiconductor
00:01:41
value chain and you can see that from the
semiconductor cheat sheet I created for you
00:01:48
you can download it for free it will be linked
in the description below enjoy it Synopsys and
00:01:53
Cadence tools are essential without these tools
the most advanced chips today like NVIDIA GPUs
00:02:00
or Apple a A-Silicon would not have been
possible however as technology scales with
00:02:05
chip makers like NVIDIA and AMD now working on 2nm
00:02:10
and even 16-angstrom designs the complexity
of chip layouts is increasing for the most
00:02:16
advanced process nodes the placement interconnect
is getting even more complex and we are facing new
00:02:22
thermal challenges power delivery becoming more
problematic and solving all of these issues takes
00:02:28
a lot of time and iterations and sometimes it can
take many weeks or even months the main problem
00:02:35
here that when we are talking about placing for
example 100 million cells on a tiny area first
00:02:41
we need to evaluate a huge number of possible
placement options to find the optimum one and
00:02:47
when we look at it as at a game chip design game
is very complex it exceeds 10 in the power of
00:02:54
billions or even trillions possible configurations
so how we can quickly find the best one to solve
00:03:02
this riddle Google introduced AlphaChip an AI tool
designed to accelerate and optimize chip layouts
00:03:09
here they are applying reinforcement learning
to chip design and their new paper highlights
00:03:15
the success of this approach it turns out that
AlphaChip can explore these huge design spaces
00:03:21
much faster than a human designer can no surprise
right and even faster than EDA tools now let's
00:03:29
dive deep into how it works similar to AlphaGo and
AlphaZero that mastered the game of Go and chess
00:03:36
AlphaChip approaches chip floor planning stage as
a kind of a game essentially chip floor planning
00:03:43
is framed as a sequential decision-making game
the game starts with an empty grid representing
00:03:50
the chip area during the game the agent places one
block after another and when it's done placing all
00:03:56
the cells and blocks it then rewarded based
on the quality of this placement and here it
00:04:02
takes into consideration different metrics like
the length of the wire interconnect we discussed
00:04:08
before and shorter is better than area performance
and power and based on this metric the AI is being
00:04:16
rewarded for good layouts and getting penalties
for the suboptimal ones and it improves through
00:04:24
practice by designing thousands of layouts it's
been trained on tens of thousands of layouts
00:04:31
and it's getting better at each iterations as we
humans do and already now it's being used across
00:04:37
different industries from data center chips
to mobile chips and this helps to reduce time
00:04:44
to market and also costs but before we deep dive
into the results and the impact of this research
00:04:51
this video is brought to you by Skillshare
Skillshare is the largest online learning
00:04:57
community offering thousands of classes led by
industry experts these classes range from computer
00:05:04
science programming and electrical engineering
to AI Innovation and business for me fall is a
00:05:10
season of new beginnings where I feel especially
motivated to focus on learning and I'm using
00:05:15
Skillshare to work on my business management
and communication skills as I lead several
00:05:21
companies now I need a better understanding on
how financial reporting works so I'm taking a
00:05:27
class on accounting led by Matt Cooper CEO of
Skillshare and it's been really helpful for me
00:05:33
I'm also working on improving my communication
skills which are so essential for both personal
00:05:39
and professional growth and I'm taking a class on
communication skills led by professor Alex Lyon
00:05:46
and I love it and I wish everyone could watch it
he explains how to become a more clear and more
00:05:53
concise communicator and how to develop that crisp
and confident sound that sets great leaders apart
00:06:00
I will leave a link in the description below make
sure to check it out the first 500 people to use
00:06:05
my link will receive a one month free trial of
Skillshare computer chips have fuelled remarkable
00:06:13
progress in artificial intelligence and now AI
wants to return the favour by making better chips
00:06:20
what's so interesting AlphaChip is already being
used for many real world designs starting from
00:06:26
Google's AI chip so-called TPU Tensor Processing
Unit it was used in the last 3 designs and to
00:06:34
their new Axion processor which is data center
arm-based CPU Mediatek also used AlphaChip to
00:06:43
design their 5G modem chip used in Samsung mobile
phones so you see it's really penetrating the chip
00:06:51
design industry according to the paper AlphaChip
can generate better computer chip designs in just
00:06:57
a few hours process that used to take humans
weeks or even months and it's managed to
00:07:04
reduce wire length by 6% compared to the designs
done by human experts shorter wire length means
00:07:12
more compact designs smaller form factor and also
faster signal propagation means faster performing
00:07:20
chips on this graph you can see the overall trend
that is getting better and better at it through
00:07:25
continual practice it's important to keep in mind
that AlphaChip focuses on the layout optimization
00:07:32
phase which is critical but very small substep
in the flow of implementing circuits into layout
00:07:40
this is just to give you a feeling that this
is just one of the small substeps in the chip
00:07:46
design flow and AlphaChip is not able to design
a chip from scratch not even close to be honest
00:07:53
we are so far from that that the light from the
finish line hasn't reached us yet what I can tell
00:07:59
you for sure there is a lot of potential here and
what's beautiful DeepMind team open sourced this
00:08:07
approach and made it available to everyone across
the community and this is sparking an entire new
00:08:13
wave of innovation beyond layout in the RTL coding
synthesis timing sign off all the other stages and
00:08:22
beyond when I attended the Hot Chip conference
at Stanford University there were talks about
00:08:29
AI chips and then AI in chip design with a looming
question will AI elevate or replace Hardware
00:08:38
Engineers and the answer is both in general we
can break it down to 2 main approaches of AI being
00:08:46
used in chip design LLM-based and reinforcement
learning based the first one LLM-based is
00:08:53
leveraging large language models which are trained
on a vast amount of design data and documentation
00:09:00
it's basically using natural language processing
to understand design requirements and constraints
00:09:06
and eventually it's able to generate RTL code
verification test benches and just assist a
00:09:12
human designer throughout the chip design process
NVIDIA for example uses LLMs to assist engineers
00:09:19
with answering technical questions debugging
design issues and more they've also deployed
00:09:25
AI agents for tasks like timing optimization
report analysis and layout generation and the
00:09:32
second example is reinforcement learning based
AI the one we discussed today AlphaChip is a
00:09:39
prominent example of it this approach treats chip
design as a complex optimization problem where
00:09:45
AI agent learns through trial and error the major
EDA players are already bringing such AI features
00:09:53
to EDA tools Synopsys for example already have
a tool with similar capabilities to AlphaChip
00:09:59
it's called DSO.ai Design Space Optimization AI
and from my conversation with Synopsys executives
00:10:07
this tool has already been used in more than
thousand productive chip design tape-outs
00:10:13
and in some cases it helped to shorten design
cycle from 2 years to just 1 year this is
00:10:20
impressive and when we look at it these both
methods LLM-based and reinforcement learning
00:10:26
based they kind of serve the same goal either
speed up time to market speed up the design time
00:10:34
or use the same amount of time but come up with a
better design and this has significant potential
00:10:40
with time it will penetrate more and more phases
of chip design and eventually enable end-to-end
00:10:46
co-optimization of hardware software and machine
learning models it's an exciting time we live in
00:10:53
let me know what you think in the comments I love
to read your comments and remember to check out
00:10:59
the cheat sheet on the semiconductor value
chain which I prepared for you to celebrate
00:11:05
200K subscribers on this channel thank you
very much for being a part of this community
00:11:11
this means the world to me and if you are not
subscribed yet consider subscribing on this
00:11:16
channel I break down the complexity and showcase
the beauty of semiconductors and AI and underlying
00:11:24
technologies and if you know someone who is eager
to stay up to date with the latest advancements in
00:11:30
tech share this channel thank you very much
and I will see you in the next video ciao
