Introducing FuriosaAI: The South Korean Pioneers
Energy-efficient South Korean AI chip company FuriosaAI has made its debut in Europe. On Tuesday, the chip firm announced that it has commenced deploying its RNGD — pronounced “renegade” — series of AI accelerators at colocation leader Equinix’s LS2 datacenter in Lisbon, Portugal. Established by June Paik and Hanjoon Kim in 2017, prior to the LLM boom, Furiosa has primarily concentrated on the South Korean market, securing achievements with prominent names such as LG Electronics. Now the company aims to introduce its RNGD-based inference platform to Europe, targeting the increasing demand for sovereign AI computing.
Furiosa’s Current Offerings
Last year, we examined Furiosa’s tensor contraction processor (TCP) architecture. Fabricated using TSMC’s 5 nm process technology, RNGD is relatively modest compared to graphics chips from Nvidia or AMD. Each PCIe card comes with 48 GB of HBM3, a memory bandwidth of 1.5 TB/s, and, according to Furiosa, can deliver 512 teraFLOPS of dense FP8 performance.
The Energy Advantage
The primary advantage of these components is that, in contrast to competitors, they are not excessively power-intensive. Each card has a thermal design power (TDP) of 180 watts. By comparison, the RTX Pro 6000, Nvidia’s nearest rival, provides double the memory capacity and processing power, with a similar memory bandwidth, while consuming 3.33 times the energy.
Furiosa’s Heavy Hitters
Eight of these accelerators constitute Furiosa’s NXT RNGD Server, a 3 kW system that features up to 384 GB of HBM, sufficient to operate relatively large enterprise models such as OpenAI’s gpt-oss 120B, LG’s Exaone 236B, or Qwen 3-30B-A3B with substantial context sizes and concurrency. Moreover, since the systems are air-cooled, they can be installed in current datacenter racks.
The European Expansion: What’s Under the Surface?
Furiosa’s entry into the European market might hinge more on brand visibility and software adaptability than merely offloading surplus RNGD accelerator stock. As previously noted, Furiosa is collaborating with Broadcom on a third-generation AI accelerator.
Innovations in the Third Generation
This partnership will enable the two companies to modify Furiosa’s Tensor Contraction Processor technology into a multi-die system-on-package that incorporates faster HBM4 or HBM4e memory. The forthcoming chip will also utilize Broadcom’s Ethernet and PCIe switching technology to facilitate larger scale-up clusters than the eight-way systems Furiosa is already developing.
Partnerships and Future Prospects
Furiosa is among various companies licensing technologies from Broadcom to bolster its next-generation accelerators. Earlier this year, Meta revealed its latest generation of MTIA chips developed with Broadcom’s assistance. OpenAI and Google have also announced chip partnerships with Broadcom, though specifics remain sparse.
Await the Next Generation!
Furiosa’s third-generation accelerators are expected to deliver significantly improved performance and scalability, but their dependence on HBM4 and HBM4e memory, which are just entering the market this year and next, indicates that we likely won’t encounter them in the wild in the near future.
Conclusion: New Chips on the Horizon! Europe Welcomes Fresh Allies
So there it is, FuriosaAI is embarking on its European venture with its moderately powered yet seemingly effective RNGD chips. Keep your eyes peeled, Europe, they’re here to stir things up, one datacenter at a time!
