AI Compute Core NPU

Compute Granule that Brings
Maximum Flexibility and Energy Efficiency


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.


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



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.

  • Number of ION™s
  • TOPS
  • Average Power
  • DL Inference throughput for DeepLOB*
    12.5K Symbols per second

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.

  • Number of ION™s
  • TOPS
  • Average Power
  • DL Inference throughput for DeepLOB*
    50K Symbols per second

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.

  • Number of ION™s
  • TOPS
  • Average Power
  • DL Inference throughput for DeepLOB*
    400K Symbols per second

System specs

TSMC 7nm
Package size
8.7mm x 8.7mm
Compute cores
Peak FP16 Perf.
Peak INT8 Perf.
Peak INT4 Perf.
2 - 6 Watt (Configuarable)
Mixed precision
Customized ISA
Various vision and language models (CNN, LSTM, BERT, etc.)

See also...


Redefining Datacenter Efficiency with Accelerative AI

Following the triumphant reception of our inaugural product, we've expanded our horizons and developed a cutting-edge datacenter inference AI chip, essential for businesses and cloud service providers. Backed by impressive global benchmark(MLperf™ v3.0) results, this product addresses the critical needs of the datacenter industry. Starting from 2023, our innovative solution accelerates transformer models with unparalleled latency, paving the way for datacenter commercialization on a global scale.
Explore ATOM™