Furiosa RNGD

The FuriosaAI RNGD, pronounced "renegade", is a South Korean inference accelerator built around a Tensor Contraction Processor rather than a conventional matrix pipeline, and its defining characteristic is power: 512 TFLOPS of FP8 inside a 180 W passively cooled dual-slot PCIe card. Eight processing elements at 1.0 GHz deliver 256 TFLOPS BF16, 512 TFLOPS FP8, 512 TOPS INT8 and 1,024 TOPS INT4, backed by 48 GB of HBM3 at 1.5 TB/s and 256 MB of on-chip SRAM with 384 TB/s of on-chip bandwidth. Built on TSMC 5nm with a PCIe Gen5 x16 interface, a 12VHPWR connector, ECC, SR-IOV and eight-way multi-instance partitioning. The low power budget is the point of the design: Furiosa fits eight cards into a 3 kW air-cooled 4U server and five of those servers into a standard 15 kW rack, where a comparable GPU rack needs liquid cooling. Furiosa labels none of its figures dense or sparse and publishes no sparse rate, so nothing here is derived. Every specification comes from the Technical Specifications table of Furiosa's own datasheet, and each one cross-checks exactly against the eight-card server specification in the same document.