FuriosaAI Brings Its AI RNGD Accelerators to Europe: A Challenge to NVIDIA in the Name of Energy Efficiency
FuriosaAI Brings Its AI RNGD Accelerators to Europe: A Challenge to NVIDIA in the Name of Energy Efficiency
The South Korean company FuriosaAI makes a significant step in its international expansion by announcing the availability of RNGD servers in Europe. The company has started installing systems at the Equinix LS2 data center in Lisbon, turning it into a hub for European companies interested in exploring an alternative to NVIDIA GPUs for AI inference workloads.
The choice of Portugal is not coincidental. In addition to the new commercial hub, FuriosaAI already has a research and development center in the Portuguese capital focused on software compilers and chip and PCB design. This presence will enable the company to provide technical support to its European customers and partners. The announcement also coincides with the RAISE Summit in Paris, an event dedicated to artificial intelligence where the company will present its hardware roadmap and developments of its proprietary software stack.
FuriosaAI's proposal stands out for its different approach compared to the leading GPU manufacturers. Rather than pursuing absolute performance supremacy, the company aims to maximize the performance-to-energy consumption ratio, an increasingly important aspect for European data centers, where the availability of electricity and network capacity often represent limiting factors.
The RNGD accelerator—already selected by LG—is manufactured using a 5-nanometer process at TSMC and utilizes SK hynix memory. At its core is the proprietary Tensor Contraction Processor (TCP) architecture, developed specifically for inference workloads. Each accelerator offers a computational power of 512 TFLOPS in FP8 precision while maintaining a TDP of 180 watts, which is significantly lower than that of many GPUs.
Eight accelerators can be integrated within the NXT RNGD server, a platform of about 3 kW designed to be installed in traditional air-cooled racks, without requiring liquid cooling systems or modifications to the existing infrastructure. According to FuriosaAI, this feature allows businesses to quickly implement platforms dedicated to large language models and agentic AI applications, avoiding the time necessary for facility upgrades.
The installation at Equinix LS2 will initially serve as a demonstration environment. European companies will be able to conduct direct tests on the performance of the RNGD architecture with state-of-the-art AI models, verifying execution speed, ease of integration, and behavior under real workloads. Additionally, a proprietary software stack with a dedicated SDK is available to support the hardware, designed to reduce the need for manual optimization of kernels and speed up the deployment of PyTorch-based applications.
FuriosaAI also highlights that its software environment represents an alternative to NVIDIA's CUDA ecosystem. The SDK uses a general compiler capable of automatically translating PyTorch code to the TCP architecture, while a Virtual ISA allows developers to intervene at a low level where maximum performance is required without facing the typical complexities of traditional GPU programming.
In addition to its European expansion, the company recently announced a strategic partnership with Broadcom for the development of the third generation of AI accelerators. The project will combine the TCP architecture with Broadcom’s interconnection and networking technologies to create a platform aimed at large hyperscalers and AI models with over one trillion parameters.
The new solution will adopt HBM4//HBM4E memory, a 2-nanometer manufacturing process, and a high-speed interconnection network optimized for modern Mixture-of-Experts (MoE) models, where communication efficiency between accelerators is one of the main performance factors. According to both companies, the goal is to improve not only computational power but especially efficiency in data exchange and performance-per-watt ratio, elements that are increasingly central to the evolution of AI-dedicated data centers.
FuriosaAI emphasizes that the RNGD is already in mass production and is being utilized in production environments by clients such as Samsung SDS and LG AI Research, while the collaboration with Broadcom represents the next step in the technological roadmap.