High-PERFORMANCE COMPUTING AT THE EDGE
NEXON: HIGH-PERFORMANCE edge ai SERVERs
Introducing the NEXON series of edge AI servers dedicated to high-performance computing (HPC) in a compact, low-power form factor.
The NEXON servers turn Big Data into tiny actionable representations allowing large-scale AI inferences ultra-quickly.
New data can be processed as it is acquired and directly augments the active instances, essentially performing AI training on the go.
The entry-level NEXON-1 server features 1200 TOPS-equivalent of AI performance in a lightweight 1U chassis. It serves as dedicated environment and perfect testbed for developers investigating the technical challenges of making HPC mobile.
The NEXON servers are ITAR-free, offer power-efficient solutions and computing flexibility with Intel i9 (default), Intel Xeon or ARMv8 processors to handle any pre/postprocessing and interface tasks.
The computing capacity scales seamlessly by adding more AI-inferencing support with NEXON-8 and NEXON-16 servers, which are 8x and 16x faster than NEXON-1, respectively. Other AI server capacity is possible and can be custom built on request.
AICRAFT can help customers quantify their Big Data computing requirements and match the edge computing solutions accordingly, to save on time, cost, storage space and energy consumption.
key benefits
technology differentiators
PARAMETER | STANDARD GPU | NEXON |
Number of classes | Limited (thousands) | Unlimited (millions) |
New sample data | Retrain model on the whole dataset (weeks)Reduced accuracy with incremental learning | Add instance to live database (milliseconds)Enhanced accuracy and extremely low latency |
Accuracy | Fixed quantisation(Prone to inaccuracy and slow) | Variable quantisation(Accurate and fast) |
Architecture | Large board, lots of RAM & storage capacity requiredBulky and costly spacecrafts | Significantly smaller RAM and more efficient boardIdeal for space to learn in-situ |
Programming skills | Specialised languages (CUDA, VHDL, Verilog)Transcoding customer code is prone to errors | Generic languages (C, C++, Python)Simple interface program to customer code |
Development time | Long | Short |
use cases
GROUND STATIONS
Smartly distribute and filter incoming signals to forward informative data to the cloud or results directly to customers, reducing irrelevant data traffic, storage and cost.
ON PREMISE
Tailored server racks and workstations for cloud-free solution or when cloud connectivity and remote communication are not possible.
ON PLATFORM
For sea surface, underwater, airborne or any situation where equipment is battery powered and requires Big Data processing.
Application examples
Synthetic Aperture Radar
- Ultra-fast beamforming and back-propagation
- 20,000 I/Q data turned into 10,000 x 10,000 image in 50 seconds! That is 100 million reconstructed pixel intensities with histogram equalisation in 50 seconds.
- Training on-the-fly in milliseconds
- Enhanced accuracy using Fast Fourier Transforms
Video processing – CCTV
Single 1U server identifies up to 2 years of 10 events recorded by CCTV cameras per second.
Synthetic aperture radar
The NEXON servers turn radar signals into an area map by applying an extremely fast back-projection method. The input is a standard CPHD (Level-0) data file from which several outputs can be produced, including the complex image, PNG or GeoTiff.
Multiple servers can be supplied in a cabinet with full software coordination and synchronisation between the servers. In the example below, the runtime performance includes the complete process of loading the input files and pulses from disk, pre-calculated digital terrain model grid and saving the complex output SAR image to disk.
Multiple servers can be supplied in a cabinet with full software coordination and synchronisation between the servers. In the example below, the runtime performance includes the complete process of loading the input files and pulses from disk, pre-calculated digital terrain model grid and saving the complex output SAR image to disk.
Input |
Total Pulses: 29,231 |
CPHD file size: 5.9 GB |
Image size: 19,569 x 19,569 pixels (382,945,761 pixels) |
Resolution: 0.204 metre |
Area size: 4,000 x 4,000 metres |
nexon-8
Server card # | Time (sec) |
1 | 96.5 |
2 | 66.5 |
4 | 39.9 |
8 | 33.5 |
16 | - |
Nexon-16
Server card # | Time (sec) |
1 | 87.1 |
2 | 67.7 |
4 | 41.2 |
8 | 33.6 |
16 | 28.6 |
HOW WE SUPPORT customers
- Quantify the requisite performance for customer applications
- Supply custom hardware and interface programs to reuse existing customer code
- Scale the solution to any size dataset or performance requests
- Support seamless deployment and integration on premise and/or on platform
- Provide tailored customer training to maximise their business operations
- Offer tailored design services for software, firmware, web/mobile applications.