ION™
AI Compute Core NPU
Compute Granule that Brings
Maximum Flexibility and Energy Efficiency
Description
Versatile compute engine amongst a plethora of NPUs
The Rebellions AI Compute Core ION™ provides the flexible inference capability with low power, a small footprint, and high performance for the edge computing systems. Featuring a customized instruction set architecture (ISA) crafted for >1K multiply-and-accumulate (MAC) units, the ION™ brings high-performance inference acceleration with exceptionally high utilization ratio compared to other AI accelerators (GPU, NPU, etc.). ION™’s versatility, compact size, and low power suffice the demands for edge deployments as in mobile.
Unmatched TOPS/Watt verified in
TSMC 7nm technology
ION™ Compute Core was fabricated in TSMC 7nm technology and successfully measured its functionality and efficiency for the benchmark networks. Supporting mixed-precision (FP16, INT8/4/2) computation with up to 2GHz operating frequency, ION™ brings >2.0 TFLOPS/Watt for FP16-based vision tasks and >10 TOPS/Watt for INT8-based language tasks, respectively.
60x end-to-end system performance on
finance and edge computing
Powered by ION™ HW/SW full stack implementation, multi-chip integrated FPGA-board is deployed to high frequency trading solution (HFT), LightTrader™, for the Wall Street based investment banks. The LightTrader™ achieves up to 60x more AI-enabled HFT performance, instantly upgrading any HFT solutions to intelligent trading system.
Performance
Competitive Energy Efficiency for Intelligent Edge
The ION™ Compute Core offers up to 10x more performance-per-watt figure even with the comparison against the state-of-the-art mobile NPUs. This opens the chances of wide ION™ deployment to different applications, including mobile companion chips, smart cities, robots and retail.
Finance Trading
Vision
Language
Key features
ION™ Brings Breakthrough Trading Solution Architecture Innovations
* DeepLOB: Deep Convolutional Neural Networks for Limit Order Books, IEEE Transactions on Signal Processing, 2018
ION™ Compute Core
The first-generation Compute Core provides FP16-based accurate stock prediction, which utilizes 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 accelerators, ION™, and the FPGA-based conventional HFT pipeline for the short-latency-high-throughput trading solutions with a minimized symbol miss rate, which is one-and-only solution for AI-based HFT trading so far.
Ultra-low Latency x High throughput Finance AI Acceleration Server
Integrating eight LightTrader™ boards into one standard 4U server, the proposed server- level solution extends the capability of our ION™-based Finance Inference up to 0.5 Peta OPS and 3.2 Tbps symbol processing throughput.
System specs
Customized ISA
Various vision and language models (CNN, LSTM, BERT, etc.)
See also...
ATOM™
