The Lacelli AI Server , says the company, is an "AI Acceleration Subsystem" for edge inferencing, which can bring extremely high performing artificial intelligence (AI) to public and corporate data center operations. The system can incorporate audio, visual, or sensor inputs to execute artificial intelligence for a diverse range of applications:
- Object detection, classification & identification
- Visual analysis
- Natural language processing
- Voice command, recognitions & authentication
- Face detection & recognition
- Image style transfer & super resolution
- Image & video search, encoding, captioning, segmentation & enhancement
The server backplane includes five Arm-based intelligent switch processors and onboard DDRs within a 2U, 19-inch rack mountable chassis that includes a twin type fan for cooling. The processors provide efficient peer-to-peer links between the AI modules within the chassis, as well coordinate communication between the server and external clusters of servers in the same data center.
The server chassis can accommodate up to 32 AI Modules, each module containing four of the company's Lightspeeur 2803 accelerator chips mounted on PCIe Gen2 x4 cards paired with an Arm-based core and having a dedicated own slot and port. Each AI Module delivers 1,200 FPS while drawing only 5 W for a performance-to-energy usage ratio of 240 FPS/W. Fully loaded, the server would deliver 38,400 FPS, with only 1.8 mS of latency.
The Lacelli Server, says the company, is ideal for many use cases:
- Machine vision
- Unmanned stores & warehouses
- Smart cities & buildings
- Business intelligence
- Artificial intelligence of things
"With this new server product, we truly open the door for many new kinds of customers," says Dr. Yasuo Nishiguchi, Chairman and Chief Executive Officer at Gyrfalcon Technology Japan. "Collaborating within our ecosystem we can achieve new levels of performance, and continue to deliver with low energy use for many types of datacenter opportunities."
The Lacelli server was developed by GTI Japan, which the company says is already being tested by tier one operators and service providers.