ION™
Scalable AI Compute Core
Compute Granule with Maximum Flexibility and Efficiency
Description
Versatile and Powerful AI Compute Engine
Rebellions' AI Compute Core, ION™, provides flexible inference capabilities with low power consumption, a small footprint, and high performance for edge computing systems. Featuring a customized instruction set architecture (ISA) designed for over 1,000 multiply-and-accumulate (MAC) units, ION™ delivers inference acceleration with an exceptionally high utilization rate compared to other AI accelerators. ION™’s versatility, compact size, and low power requirements make it the optimal choice for edge deployments.
Energy Efficiency with TSMC 7nm Technology
The ION™ Compute Core was fabricated using TSMC’s 7nm process and successfully demonstrated its functionality and efficiency on benchmark networks. Supporting mixed-precision (FP16, INT8/4/2) computation with up to 2GHz operating frequency, ION™ delivers over 2.0 TFLOPS/Watt for FP16-based vision tasks and over 10 TOPS/Watt for INT8-based language tasks, respectively.
60x Performance for Finance and Edge Computing
Powered by ION™’s hardware and software full-stack implementation, LightTrader™, the Field Programmable Gate Array (FPGA) integrating multiple chips, has been deployed in high-frequency trading (HFT) solutions at Wall Street-based investment banks. LightTrader™ achieves up to 60x AI-enabled HFT performance, instantly upgrading any HFT solution to a more intelligent trading system.
Performance
Competitive Energy Efficiency for Intelligent Edge
ION™ achieves up to 10x more performance-per-watt figure even in comparison against the state-of-the-art mobile Neural Processing Units (NPUs). This makes ION™ ideal for various applications including mobile companion chips, smart cities, robots and retail.
Finance Trading
Vision
Language
Key features
ION™ Brings Innovative Trading Solution Architecture
* DeepLOB: Deep Convolutional Neural Networks for Limit Order Books, IEEE Transactions on Signal Processing, 2018
ION™ Compute Core
The first-generation Compute Core ION™ supports FP16-based accurate stock prediction, utilizing the single batch inference and predicative execution pipelines. ION™ also supports BFloat16 and low precision integer operations such as INT8/4/2.
LightTrader™ : World-first AI-enabled HFT Card
This PCIe card integrates the custom AI accelerator, ION™, and the conventional FPGA-based HFT pipeline for low-latency and high-throughput trading solutions with minimized symbol miss rate. It is the one-and-only solution for AI-based HFT trading so far.
Ultra-low Latency, High throughput Finance AI Acceleration Server
Integrating eight LightTrader™ boards into one standard 4U server, the proposed server- level solution extends the capability of the ION™-based finance inference up to 0.5 PetaOPS and 3.2 Tb/s throughput for symbol processing.
System specs
Customized ISA
Various vision and language models (CNN, LSTM, BERT, etc.)
See also...
System Solution
Start Lean, Scale Green
ATOM™
Inference AI Accelerator for Data Centers
With international recognition from the industry-standard benchmark MLperf™ v3.0, ATOM™ can be scaled to accelerate state-of-the-art AI models of various sizes.
