Cognifiber benchmarks its photonic AI performance
Israeli startup CogniFiber has benchmarked the performance of its photonic AI technology as part of its product integration tests.
CogniFiber’s approach implements direct analog neuromorphic photonic computing in an optical fibre, removing all bottlenecks from AI inference so that the task speed is limited only by the input clock.
The system is expected to reach speeds of 100 million tasks per second, surpassing other silicon-based AI accelerators such as Nvidia’s DGX A100 estimated 5 million tasks per second or Lightmatter’s Envise Server with 24 million tasks per second.
The photonic AI processing is also more power efficient, at 350,000 tasks per watt.
Related articles
- Cognifiber raises $6m for photonic AI processing
- Glass processors aim to boost edge AI
- Photonic AI chip for fast classification
CogniFiber is developing a full solution for data cetnre AI applications based on its proprietary fibre-based technology together with standard optical communication devices. This requires a fraction of the space to achieve similar computing power, while significantly reducing cooling and operational overhead.
“Our initial findings are outstanding, displaying the capabilities to support the mega data centres of tomorrow,” said Dr. Eyal Cohen, Co-founder & CEO of CogniFiber (above). “Operating at room temperature without dissipating significant heat to its surroundings means data centres can offer customers greater uptime reliability with lower cost.”
“Our new approach to developing and harnessing large data processing capabilities allows for companies to begin bringing AI and Machine Learning capabilities to the edge of their networks,” said Professor Zeev Zalevsky, Co-founder & CTO of Cognifiber.
Other articles on eeNews Europe
- Configurable image signal processor core for IoT smart camera chips
- First PowerPC SoSa VPX board
- €140m 3D chip research centre opens in Dresden
