ATOM

ATOM is a neural processing unit (NPU) from Rebellions, a South Korean AI chip company founded in 2020 and valued at roughly $1.4 billion as of late 2025. It is purpose-built for large-scale AI inference, targeting developers and enterprises that need high throughput and low latency for serving large language models (LLMs) and small-to-medium enterprise applications. The hardware delivers 128 TFLOPS FP16 and up to 1024 TOPS INT4, with 64GB of ultra-low latency memory, making it suitable for high-concurrency, latency-sensitive environments. ATOM is designed for data scientists, ML engineers, and DevOps teams who want a GPU-like software experience but with the efficiency and cost benefits of a specialized accelerator. It is particularly relevant for organizations deploying LLMs at scale, such as chatbots, code assistants, and enterprise search, where performance-per-dollar is a key metric.

The core of ATOM's software stack is the RBLN SDK, which provides native integration with PyTorch, vLLM, and Triton. Developers can use standard PyTorch APIs, with graph-mode optimization for hardware-accelerated execution and eager mode for flexible experimentation. The SDK supports intelligent precision control across FP32, FP16, FP8, FP6, and FP4, allowing users to balance accuracy and performance. Distributed inference is handled via the Collective Communication Library, which manages tensor parallelism across multiple devices. The RBLN Compiler automates model compilation and distribution, while built-in profiling and debugging tools help identify bottlenecks. For LLM serving, vLLM-RBLN provides a high-throughput, low-latency backend that is production-proven. The system also integrates with Hugging Face for fast model onboarding and supports custom kernel programming via the Triton backend.

ATOM competes directly with GPU-based inference solutions from NVIDIA (e.g., H100, B200) and other NPU/ASIC offerings from companies like Groq and Cerebras. However, its primary differentiator is performance-per-dollar, especially for small language model (SLM) applications. Rebellions claims that ATOM delivers an unbeatable performance-per-dollar ratio, validated by MLPerf benchmark results, and is engineered for low power usage. The company is South Korea's first AI chip unicorn, having raised over $540 million in funding. While it does not directly compete with text editors like Atom (which is discontinued) or diagramming tools like Miro and Creately, those were listed as competitors in the source data, likely due to a naming collision. In the AI accelerator space, its main rivals are NVIDIA GPUs and other inference-focused chips.

ATOM's main trade-offs involve its software maturity and ecosystem depth. While the SDK is designed for a GPU-like experience, some features require technical knowledge, and documentation has been noted as a pain point by early users. Customer support response times vary, and the mobile app (if applicable) has limited functionality. The current usage flow of the RBLN SDK has constraints, such as requiring explicit model compilation steps, though the team is shifting toward a more PyTorch-native UX with torch.compile integration. Additionally, ATOM is optimized for inference, not training, so it cannot replace GPUs for model development. Finally, as a relatively new entrant, its long-term support and community ecosystem are less established than NVIDIA's CUDA ecosystem.

Contact for pricing; no public tiers or per-unit costs are disclosed.