5nm Versatile Inference SoC
Versatile yet Energy Efficient
AI System-on-Chip
Standing on par with industry’s leading competitors, ATOM™ delivers top performance across various types of AI tasks like computer vision, natural language processing and recommendation models. Utilizing the silicon-proven neural core ION™ as the compute granule that scales up with perfect linearity, ATOM™ is the optimal chip for large-scale inference required in edge computing and data centers.
Scalable, energy-efficient, and fast: ATOM™ is poised to be a pivotal enabler for server-level AI services, optimizing the AI-as-a-Service (AIaaS) stack. As a multi-core System-on-Chip (SoC), ATOM™ is built upon a dataflow architecture to deliver superior inference performance. Supported by SR-IOV-based user-level parallelization, our proprietary multi-level dependency control mechanism minimizes latency overhead for inter- and intra-chip data orchestrations. ATOM™’s high system utilization rate is enhanced by up to 40% through the seamless interplay of our hardware, software, and firmware.
Samsung Foundry’s advanced EUV process node enables ATOM™ to deliver improved performance, reduced power consumption and enhanced energy efficiency. Geared with PCIe Gen5, GDDR6 and high-speed I/O, ATOM™ is the optimal AI accelerator to serve different markets, spanning from edge computer to data centers.
Inference v3.0, Single Stream
At MLPerf™, ATOM™ showed impressive results across both vision and language models, outperforming Qualcomm and NVIDIA, with 0.24ms in ResNet50 Single Stream and 4.30ms in BERT-Large Single Stream, respectively.
* Shorter the faster
* Shorter the faster
ATOM™ is built on the small yet powerful silicon-proven AI core ION™, which, with its Coarse-Grained Reconfigurable Architecture (CGRA), is flexible, programmable, and scalable. This design allows ATOM™ to handle networks with varying depths and complexities, from small applications to hyperscale services, all while maximizing energy efficiency.
ATOM™ is a multi-core System-on-Chip, consolidating all essential components onto a unified substrate. The communication overhead within the chip is masterfully handled via on-chip data dependency handling.
ATOM™ can accelerate different types of neural networks, including Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM), Bidirectional Encoder Representations from Transformers (BERT) and recent transformers-based models (e.g. T5, GPTs, etc.).
ATOM™ can be partitioned to support up to 16 separate jobs simultaneously, with each job isolated at both the hardware and software levels. ATOM™’s Multi-Instance feature allows higher resource usage and flexibility, optimizing utilization and performance.
System Solution
ION™
