Gyrfalcon launches GAINBOARD 2803S inference accelerator optimized for performance-per-watt efficiency.

Gyrfalcon Technology Inc. (GTI) announced on November 29, 2018, its second-generation server solution, the GAINBOARDTM 2803S Inference Accelerator Server. The system can be installed as an upgrade to existing data centers to rapidly deploy AI acceleration for large numbers of data models running in parallel. Providing 1,084 TOPS at 110W, the GAINBOARDTM 2803S offers the industry-best ratio of performance-to-power use and low cost of ownership for data centers.

Unlike general-purpose GPUs designed to run fast but consume significant power, GTI's server solution has been built from the ground up to process AI data within the constraints of data center environments. Frank Lin, president of GTI, stated: "Many data center operators have complained about having to wait for product from other providers, and now they can quickly get the solution they need along with increased performance and significantly lower power use." He further emphasized that the solution addresses data center operators' primary concerns: "In the data center, the top business concerns are the ability to provide higher service than competitors while keeping costs down. Our GAINBOARDTM 2803S Inference Accelerator Server has a lower equipment cost than competing solutions, delivers a lower monthly energy bill and reduces cooling equipment requirements. It is an all-around win to achieve AI acceleration without the drawbacks that data center providers have experienced with other AI solutions."

Each GAINBOARDTM 2803S Inference Accelerator Server contains dual Xeon processors and 64 individual Lightspeeur® 2803S chips housed in four PCIe cards, with each card containing 16 chips supported by four Xilinx FPGAs. At its core is GTI's proprietary Matrix Processing Engine (MPE) and AI Processing in Memory (APiM) architecture, which accelerates AI using convolutional neural networks (CNN). A single Lightspeeur® 2803S chip delivers 16.8 TOPS at just 0.7W with latency as low as 2 milliseconds, supporting ResNet, MobileNet, ShiftNet, and VGG neural networks for inference and training.

The design allows AI models to be processed in parallel and operate in cascade mode, where large and complex models can be spread across individual chips without requiring host processor intervention, maintaining high performance and low energy use. The GAINBOARDTM 2803S Inference Accelerator Server will be available from early Q1 2019, joining existing GTI solutions including server configurations using the 2801S multichip board for advanced edge and data center applications, and the Lightspeeur® 2801S and 2803S Neural Accelerators. Products are already shipping commercially to electronics companies including Fujitsu, LG, and Samsung, with availability for qualified customers now.