The chip industry in 2024 ushered in a dazzling new star, FuriosaAI's AI accelerator RNGD ("Renegade"), which was unveiled at Hot Chips, heralding a disruptive change in the data center AI hardware landscape. Founded in 2017 by former AMD, Qualcomm, and Samsung engineers, the company has successfully brought RNGD to the front of the market with its strategy of rapid innovation and product delivery.
Breakthrough RNGD accelerator
The close collaboration between FuriosaAI and TSMC enabled RNGD to launch quickly after receiving the first silicon samples, which once again confirmed FuriosaAI's agility and efficiency in technology development. Back in 2021, just 3 weeks after the launch of its first-generation chip, FuriosaAI submitted an astonishing score on the MLPerf benchmark and achieved a whopping 113% performance leap with compiler enhancements in subsequent submissions.
Outstanding performance and efficiency
RNGD's early test results have been impressive, and it has performed well with large language models such as GPT-J and Llama 3.1. For models with about 10 billion parameters, a single RNGD PCIe card is capable of delivering 2,000 to 3,000 tokens per second, a performance that is unique in its class.
Innovation leads, continuous progress
June Paik, co-founder and CEO of FuriosaAI, couldn't be prouder of the launch of RNGD: "RNGD is the culmination of years of innovation that not only achieved single-use silicon success, but also had an exceptionally fast start-up process. Paik emphasized, "RNGD provides a sustainable and accessible AI computing solution that meets the industry's real-world needs for inference."
Figure: FuriosaAI launches AI inference chip
The best of both worlds: performance and efficiency
RNGD's innovation lies in its non-matrix multiplication-based tensor shrink processor (TCP) architecture, which offers the perfect balance between efficiency, programmability, and performance. With a powerful compiler, RNGD treats the entire model as a single fusion operation, further optimizing performance. At the same time, RNGD's TDP is only 150W, which is undoubtedly more attractive than leading GPUs that consume more than 1000W.
The response from the industry has been overwhelming
Vik Malyala, Supermicro's senior vice president of technology and artificial intelligence, praised the Furiosa RNGD AI inference solution as driving the adoption of Supermicro's green computing. By integrating Furiosa's technology, Supermicro systems reduce power consumption per card while still delivering superior inference performance.
Aditya Raina, Chief Marketing Officer of GUC, also expressed high recognition of FuriosaAI's partnership, citing RNGD's success as a result of a deep understanding of modern AI software and hardware, as well as FuriosaAI's outstanding performance from design to delivery.
RNGD is currently sampling to early access customers and expects to be generally available in early 2025. With FuriosaAI's presentation and live demo at Hot Chips, we have reason to believe that RNGD will open a new chapter in the AI accelerator space, bringing unprecedented performance and efficiency to the data center.
