Jetraw Core — Improve data scalability & reliability with raw image compression

Store more & manage data better
Improve accuracy of data
Decrease infrastructure & cloud costs
100PB
original raw dataset
16PB
compressed raw dataset
Compression ratio
From 5:1 to 10:1
Compression speed
FPGA
200Mpx/s/core
Software
200MB/s/core
Data quality
no artifacts
no bias
no filtering
Coming to AutoSens September 19—21? Visit our booth #49 to see raw image compression in action

Large image data in the way of automotive computing efficiency

ADS data growth

With the race towards L4+ vehicles, increased number of sensors per car, resolutions, frame rates, etc., the data produced by the systems grows exponentially. So are the costs.

25-60/car
Number of sensors in L4 vehicle
0.4TB/h → 19TB/h
Data rate growth produced by 1 autonomous vehicle
175 zettabytes
Projected data generation from connected cars by 2025
$0.6-10M/year
Costs for 1 vehicle data storage & access on AWS S3

Thus, efficient and cost-effective data handling becomes crucial. Existing infrastructure cannot cope with the demand.

PROBLEM

Traditional lossless compression struggles with scalability due to limited speed and efficiency. While visually lossless algorithms compromise the reliability of ML models for the safety of connected and autonomous vehicle’s (CAV).

Jetraw Core — real-time compression with immediate benefits to your data

WHAT IS iT

Jetraw Core is a high-efficiency low-latency 5-10:1 raw image compression integrated into certified environments, either on-board a vehicle system or in a data centre.

TURN DATA INTO MEASUREMENT

Jetraw applies lossless compression to the signal, following lossy compression to the noise. Thus it retains metrological properties of raw images while achieving higher ratio and speed than traditional lossless methods.

Unlike visually lossless codecs, it preserves raw quality and treats images as measurements, enabling reliable data-centric AI analysis and model training.

Two implementation routes

As a software to maximise your IT infrastructure and save on storage
On FPGA in the system to increase speed & reduce compute requirements

JETRAW IN A NUTSHELL

High speed
High ratio
Raw quality
No artifacts
Tailored for AI

Accelerate your ADS time-to-market with 5-10x more seamless data management

Faster acquisition & I/O

Acquire and transmit data faster between memory, FPGA, CPU, on-board, remote and cloud storage, using existing cable and network

Higher throughput

Achieve low-latency higher data throughput rate (FPS, MB/s, Mpx/s), while saving computational and bandwidth resources

Re-use infrastructure & lower costs

Increase storage capacity and generate 5-7x more data on test vehicles without upgrade. Reduce data offloading frequency & storage swaps. Reduce operational & storage costs

More scalable cloud

Upload/download faster from cloud, manage your data more efficiently and cost-effectively

Accurate data for reliable AI

Jetraw delivers format-agnostic raw output in highest quality. The better the quality of data - the higher the accuracy of ML models for safer cars.

Lower energy consumption

Reduce data-related CO2 emissions with smaller storage volume, efficient logistics, and reduced power consumption from on-board to data centres
Get in touch to improve your data performance and reduce costs

Free up time and money and let the data finally work for you

The large image data produced by Level 4-5 autonomous vehicles encounters slow upload/download speeds, and the rapidly expanding volume faces recurrent system capacity limitations.

The costs of cloud storage and access could run into millions of dollars. Jetraw Core addresses such issues and unlocks scalability for OEM and Tier 1 suppliers.

SAVINGS WITH JETRAW

Original raw
Jetraw
Acquisition volume for 1 day (8h), 1 car
11TB
1.8TB
Acquisition volume for 1 year, 1 car
2.9PB
0.5PB
Time to upload daily volume over 10 Gbps
2.4h
0.4h
Time to download yearly volume over 10Gbps once from Azure Blob
27 days
3 days
Storage costs/year, incl. 1 backup
1.5M
0.2M

Jetraw Core: the first high-ratio compression designed for machine vision and ML pipelines

Sensor
On-sensor processing (HDR, companding)
Jetraw
Lossless
Typical lossy compression (e.g. JPEG2000)
Introduces bias
No
Yes, specified
No
No
Yes, unspecified
Introduces fake correlations /removes real correlations
No
No
No
No
Yes, unspecified
Reduces SNR
Yes, specified, e.g. t°
Yes, specified
Yes, specified
No
Yes, unspecified
Destroys signals below the noise level
No
No
No
No
Yes, unspecified
Compression ratio
1:1
1.4:1
6:1
2:1
6:1

Instant benefits for every petabyte of your data

Cost of cloud storage
& network, USD/year

1→0.1M

CO2 emissions,
tonnes/year

52→9t

Time required
to transfer on 100Gbps

22→4h

Integration type
FPGA IP Core
High throughput, low latency, power-efficient raw compression on FPGA
Software
Fast, easily integrated in-camera raw compression as a software
5–10:1 compression ratio
Indistinguishable from raw, interoperable
CMOS, sCMOS, CCD camera support
Mono, Bayer and other CFA image sensors support
No bias, no artifacts, no artificial correlations, no low pass filtering
Tightly-controlled image quality 1.2dB SNR equiv. increase ISO100→ISO115
Image data
  • 16-bit images
  • Configurable image dimensions
Raw image buffer or common file formats
Performance
  • 1 to 32 pixels per clock cycle
  • Up to 200 MHz clock frequency
  • Low latency (~1-line, 2-lines for Bayer)
  • 200MB /s/core
  • Multi-threading support
Integration features
  • Backpressure support
  • This is some text inside of a div block.
  • From 3968 LUT for single core compressor to 70790 for 32 pixels
  • AXI4-stream data interface
  • Available as a software library / codec
  • Callable from C, C++, C#, Java, Python
System support
  • Xilinx FPGAs
  • Intel on request
  • Intel, AMD and ARM CPU support
  • Linux, Windows and Mac support
coming soon

Future-proof your AI/ML with Jetraw algorithm unique features

Synthetic data generation for data augmentation based on physical-model
Image traceability and quality assurance to add robustness to ML models
Enhanced data normalization and testing with low compute requirements

Take next steps to maximise the value of your image data

1
Discovery
Align your requirements
with the solution
What you get
Feasibility study, your use case evaluation, performance and benefits estimation
2
Initial Implementation
Validate solution inside
your system
What you get
Performance report and a roadmap to full integration
3
Full Integration
Deliver production
version
What you get
Fully tested Jetraw Core integrated into your system