The document discusses EdgeAI, a startup developing an AI chip with a custom machine learning accelerator and new embedded memory technologies targeting low-power, high-performance edge applications. In early stages, EdgeAI aimed to enable AI vision on battery-powered cameras but faced challenges competing with GPUs. It later found product-market fit enabling solar-powered security cameras by developing a chip that performs inference with no idle power consumption. EdgeAI will validate this approach with pilots and plans to fabricate a second silicon chip and raise seed funding to develop end-to-end prototypes.
Report
Share
Report
Share
1 of 19
More Related Content
Edgeai Engr245 2021 Lessons Learned
1. EdgeAI
The future of edge computing
Mentor
Mentor
Serena
DAYAL
95 total interviews
Day One:
Offering a custom silicon chip with
new embedded memories and a
custom Machine Learning
accelerator targeting low-power,
high-throughput, and low-latency
applications.
Today:
Enable AI vision applications
on next generation battery
powered surveillance
cameras.
Massimo
GIORDANO
Kartik
PRABHU
Jo
ZHU
Weier
WAN
Alon
DROR
Picker
PhD, EE
Hacker
PhD, EE
Hustler
MBA, M.Ed
Designer
PhD, EE
Hustler
MBA
Dave
NEWMAN
2. EDGE AI - THE FIRST ALL-IN-ONE AI CHIP
● In the lab, we:
○ 2 years of research and development
○ Collaboration with a foundry that offers the new memory
○ Built the first all-in-one AI chip prototype
Our already
fabricated AI chip
Possible use cases
We envisioned
● Objective: Enabling new AI applications at the edge, with a
faster and lower power DNN accelerator enabled by new
memory technology (Resistive RAM).
Unclear purpose statement
?What does it mean?
Week 0 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7 Week 8 Week 9 The Future
Week 1
3. WEEK 1 - EdgeAI was:
● A silicon AI accelerator with new
memories for low power and high
performance inference on edge
devices
Week 0 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7 Week 8 Week 9 The Future
Week 1
EdgeAI
Security
Cameras
4. WEEK 1 - EdgeAI was:
● A silicon AI accelerator with new
memories for low power and high
performance inference on edge
devices
Week 0 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7 Week 8 Week 9 The Future
Week 1
EdgeAI
Security
Cameras
No need for AI accelerator (Not compute
intensive)
Wrong power target (0.1 mW instead of
100mW)
5. WEEK 1 - EdgeAI was:
● A silicon AI accelerator with new
memories for low power and high
performance inference on edge
devices
Week 0 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7 Week 8 Week 9 The Future
Week 1
EdgeAI
Security
Cameras
Computation power much smaller than
propulsion power
6. WEEK 1 - EdgeAI was:
● A silicon AI accelerator with new
memories for low power and high
performance inference on edge
devices
Week 0 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7 Week 8 Week 9 The Future
Week 1
EdgeAI
Security
Cameras
High-end products can afford expensive
hardware (GPUs)
Needs high compute power (4k images at
60fps) - hard to compete with NVIDIA
GPUs
Battery life is not critical
7. By WEEK 2 we got confused
Other embedded non-volatile memories offer similar performance
Competitors already use eFLASH and MRAM
eFLASH won’t
scale, but TODAY
has the same
performance of
RRAM.
MRAM offers
the same
benefits of
RRAM
We FORGOT why
RRAM was great...
Week 0 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7 Week 8 Week 9 The Future
Week 1
9. “Lost in Space”: How we reached market fit!
● Radiation resilient & affordable alternative to FPGAs for AI applications in space
almost
Huge interest from NASA engineers:
Week 0 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7 Week 8 Week 9 The Future
Week 1
Market:
● Big satellites (NASA & ESA)
● Nano satellites
“A huge
opportunity!
We can help you
make this happen.
Let’s write a grant
together”
10. “Lost in Space”: How we reached market fit!
almost
Week 0 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7 Week 8 Week 9 The Future
Week 1
● Big satellites (NASA & ESA)
○ RRAM is resistant to radiation, but is still not space graded
● Nano satellites
○ Small-market, 1000 satellite/year less than 10k whole mission
“I just launched into space an Intel’s Myriad
vision processor” - AI expert, ESA
Commercial products already sent in space
Users Customers
11. Focusing on the value proposition
AHA moment:
“Who cares about RRAM?!
What matter is what we do with it!”
Heidi Roizen
Threshold Ventures
Focus on
the value
proposition!
What is unique to RRAM,
that is otherwise NOT possible?
Week 0 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7 Week 8 Week 9 The Future
Week 1
12. Customers Need AI!
“Something that consumes 10mWh/day
would be amazing”
- Ring
“Currently I need a GPU card for
every camera and most of the time
nothing is happening”
“AI would solve a lot of problems,
but it consumes too much power”
Week 0 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7 Week 8 Week 9 The Future
Week 1
13. Searching for Optimal Product/Market Fit
● Low-power solutions can not run large complex DNNs
● High-performance compute consumes a lot of power
Low-power…
but can’t do much!
Prius
High performance…
but high power!
vs.
Week 0 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7 Week 8 Week 9 The Future
Week 1
14. Aha! RRAM Enables No Idle Power. Search for these customers!
● Great specs for AI
● No power to retain memory
● Fast wake-up
Edge AI
High-performance
when needed
10ms 1 second
Week 0 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7 Week 8 Week 9 The Future
Week 1
It goes FAST!
But no power when
stopped.
15. We Find Product/Market Fit!
● Edge AI supports 50MB DNN models, capable of waking-up and
performing 1 inference in just 1 mJ
Solar panel
provides
only
5 mW*h/day
We enable AI vision
on solar-powered
security camera
Not possible today!
Higher security,
Fewer false alarms
No need to change batteries
Happier
customers!
Week 0 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7 Week 8 Week 9 The Future
Week 1
16. MVP: smart security cameras w/o changing batteries
We are really efficient in doing this
1 frame / second
No person
Person
10 MB model
1mJ
30 frame / second
higher resolution
0.3%
Face recognition
40 MB model
5 mJ
Owner Unknown
Less Efficient but infrequent
10%
Unsafe
1000 MB model
200 mJ
safe
Above memory
available on-chip
Limited power
budget
Week 0 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7 Week 8 Week 9 The Future
Week 1
17. Aha! Software is Key! Use Existing AI IP
● Custom ML accelerator needs complex software stack (hard to accomplish)
● Instead, use an existing AI IP that offers the full software stack (Expedera)
“We’ve spent a lot of time
writing software for our current
chip, and can’t afford to redo
everything for a new chip” -
Embedded systems startup
Week 0 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7 Week 8 Week 9 The Future
Week 1
18. Quick wake-up powerful
AI acceleration that
consume no power
when idle
IP providers
Only chips: OEM
TSMC
already
collaborating
IC Design Tools
Software
Run infrequent
complex AI tasks
preserving battery life.
Engineering,
Development
Customer
Customization
Work with
customers to enable
their applications
Supply Chain
IP protection
Provide custom features
based on customer need
Physical: System
Integrator
Radiation resistant,
space-graded AI ASIC
accelerators
NASA, ESA
Sell custom-designed silicon
chips
Development: MPW runs, EDA tools, IP licenses,
Packagings, mask costs
Develop custom silicon
Contracts for AI expert
consulting for co-designing
specific applications
Develop custom
DNN accelerators
AI IP with full
software support
Enable new AI applications
on their device Product Managers
Offering new AI features
EE/CS Engineers
Development and design
Smart AI
battery
powered
Cameras
Lower-power
compute
Ease of
integration
and use
19. What’s Next:
● More product validation:
○ Further potential customers interviews
● Run pilots:
○ With various security cameras and doorbells companies
● Fabricate 2nd silicon chip:
○ Scheduled for October 2021
● Raise first seed investment in Q3-Q4 ‘21
○ Development of end-to-end prototype
Contact us: Massimo Giordano - mgiordan@stanford.edu
Week 0 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7 Week 8 Week 9 The Future
Week 1