FuriosaAI
FuriosaAI builds powerful and energy-efficient AI accelerators.
FuriosaAI has been building for this moments since 2017. - Two generations of silicon, ready as the inference era begins.
Furiosa RNGD (Renegade) : purpose-built for energy-efficient inference
- 512 TFLOPS (FP8)
- 2xHBM3 48 GB Memory Capacity (1.5 TB/s)
- 256MB SRAM (384 TB/s)
- 180W TDP targeting aired-cooled datacenters
RNGD is in mass production, ramping to 100K+ units in 2026-2027
NXT RNGD Server: 4 PFLOPS and 384 GB HBM3 in only 3kW
- 8x RNGD cards
- 4 PFLOPS total
- 384 GB HBM3 Capacity
- 12 TB/s Memory bandwidth
- 3kW Power consumption
Featured products
RNGD (Renegade)
RNGD is FuriosaAI's second-generation AI accelerator, built on the company's proprietary Tensor Contraction Processor (TCP) architecture — a design that optimizes tensor contraction operations directly, rather than relying on traditional matrix-multiplication approaches. RNGD is engineered for enterprise and cloud deployment of LLM and multimodal models at scale, targeting the demanding throughput and latency requirements of production AI services.
A defining characteristic of RNGD is its efficiency: the chip delivers high-performance LLM and multimodal inference within a radically efficient 180 W power envelope. Independent benchmarks show 1.8–2× better concurrency per kilowatt compared to Nvidia's RTX PRO 6000 across multiple service-level objectives, while maintaining interactive response times — translating into meaningfully higher inference capacity per rack for data center operators.
A defining characteristic of RNGD is its efficiency: the chip delivers high-performance LLM and multimodal inference within a radically efficient 180 W power envelope. Independent benchmarks show 1.8–2× better concurrency per kilowatt compared to Nvidia's RTX PRO 6000 across multiple service-level objectives, while maintaining interactive response times — translating into meaningfully higher inference capacity per rack for data center operators.