5nm Versatile Inference SoC
Versatile yet Energy Efficient
AI System-on-Chip
Competing against state-of-the-art AI accelerators, ATOM™ delivers uncompromised inference performance across different types of ML tasks, computer vision, natural language processing and recommendation models. ATOM™ utilizes the silicon-proven neural core, ION™, as a compute granule that scales up with perfect linearity for large- scale inference operations required in edge computers and datacenters.
ATOM™ combines ML specialized dataflow architecture and many-core SoC architecture to bring the best inference performance with SR-IOV based user-level parallelization.
In order to minimize the latency overhead for data communication and synchronization in inter-/intra- chip orchestrations, ATOM™ adopts our proprietary multi-level dependency control mechanisms and interrupt protocol that enhance overall system utilization by up to 40% of total end-to-end inference latency.
ATOM™ processor is a key enabler for enterprise/ personal server-level AI services, perfecting the customers’ AI-as-a-Service (AIaaS) stack.
ATOM™ is built with Samsung Electronics’ most advanced EUV process node, delivering the best performance and energy efficiency.
Geared with PCIe Gen5 and GDDR6 high-speed IO technologies, the ATOM™ processor can serve different markets, spanning from edge computer to datacenters.
Inference v3.0 (Latest), Single Stream Latency
ATOM™ readily supports the most advanced Al networks, including state-of-the-art GPT language models. Built with the most advanced silicon technology, ATOM™ delivers the best-in-class performance and energy efficiency.
* Shorter the faster
* Shorter the faster
Rebellions’ silicon-proven AI core delivers outstanding performance owing to its highly utilized dataflow and efficient hardware implementation.
The AI core can readily handle networks with different depths and complexities from tiny applications to datacenter-level services.
Energy efficiency is maximized by implementing a low-power design methodology and using tidy core structures.
With ION™ Core tightly orchestrated with an internal command processor, ATOM™ accelerates both large and small networks with outstanding efficiency. The overhead of extra communication is neatly hidden by the on-chip level control and data dependency handling.
ATOM™ can accelerate different types of neural networks efficiently, including convolutional neural networks (CNNs), long short-term memory (LSTM), bidirectional encoder representations from transformers (BERT) and recent transformer networks (e.g. T5, GPTs, etc.).
ATOM™ can be partitioned to support up to 16 jobs simultaneously, isolated at different HW/SW levels.
This feature provides a clear competitive advantage to service providers by boosting the user capacity and the effective usage of computing farms.
ION™
