Notifications
Clear all

Ultra-compact, low-latency ML inference on MCU, DSP, and FPGA

6 Posts
2 Users
0 Likes
638 Views
Altaf
(@altaf)
Active Member Member
Joined: 1 year ago
Posts: 4
Topic starter  

Hi. My name is Altaf Khan. I am the CEO of Infxl. Infxl develops tiny, low-latency, hardware-agnostic AI for tabular and time-series data.

On a Cortex-M4, our inference engine occupies 110 bytes. On a 200 MHz FPGA, it has a latency of 35-55 ns. You can also deploy it on any other 8/16/32/64-bit MCU, DSP, or FPGA.

Applications include IoT edge devices, predictive maintenance, healthcare/fitness wearables, high-frequency trading, and many more.

If it interests you, let us set up a quick call. Otherwise, please point me to those in your network who may find it beneficial.

You may visit removed link to test-drive our solution.

Have a great day!

Altaf


   
Quote
Laurence Liew
(@laurenceliew)
Estimable Member AI Ready Clinic Group
Joined: 2 years ago
Posts: 105
 

Thanks @altaf for sharing your company's solution. To gain better visibility - do consider submiting an actual deployed AI use case, preferably in Singapore.

 

Cheers!

Outcompute to outcompete | Growing our own timber


   
ReplyQuote
Laurence Liew
(@laurenceliew)
Estimable Member AI Ready Clinic Group
Joined: 2 years ago
Posts: 105
 

@altaf - your first post included a link, which the forum system automatically removes (As part of SPAM control). With your FIRST post now approved, you can post links to your website, and product pages etc.

Cheers!

Outcompute to outcompete | Growing our own timber


   
ReplyQuote
Altaf
(@altaf)
Active Member Member
Joined: 1 year ago
Posts: 4
Topic starter  

USE CASE #1: IoT Condition Monitoring

Infer the state of a fan based on instantaneous sensor data

 

STATES: Normal; Low voltage; Stuck object; Obstructed

SENSORS: Triaxial Accelerometer; PWM; Tachometer

 

FPGA-BASED INFERENCE:

240 LUT, 340 FF, 0 DSP on a Microchip PolarFire® 200 MHz FPGA 
Structure/parameters updateable without pausing the operation

96% accuracy with 4,000,004 samples
57 nJ/inference
18 μs/inference for the sequential implementation of Infxl inference engine


   
ReplyQuote
Altaf
(@altaf)
Active Member Member
Joined: 1 year ago
Posts: 4
Topic starter  

USE CASE #2: Particle Identification (High-Energy Physics)

Identify, in real-time, events of interest in proton-proton collisions at the CERN Large Hadron Collider

TASK: Distinguish between five elementary particles

FEATURES: 16 Physics-motivated, high-level features crafted by experts

 

FPGA-BASED INFERENCE:

35 ns/inference for the parallel implementation of the Infxl inference engine on a 200 MHz FPGA
The solution does not use DSP blocks

72.1% accuracy with 830,000 samples


   
ReplyQuote
Altaf
(@altaf)
Active Member Member
Joined: 1 year ago
Posts: 4
Topic starter  

USE CASE #3: Battery-less 60x45x2 mm Card for Gesture Recognition

Recognize gestures performed using a credit-card-like battery-less device

 

TASK: Recognize up-down and side-ways movement gestures

SENSOR: Triaxial accelerometer

 

MCU-BASED INFERENCE:

MCU-based card operates solely on the energy from the solar cell that covers one side of the card

94% accurate recognition

 

REFERENCE:

SLIDES:

VIDEO:


   
ReplyQuote
Share: