FuriosaAI, a new Korean AI chip creation, announced a cooperation with LG AI Research to jointly promote the large-scale application of the energy-efficient AI accelerator "RNGD (Renegade)." Although Nvidia's H100, B100 and other products are in line with the global data center market, LG has selected start-up manufacturers whose bets have not yet been produced in quantity. The key reason lies in "energy efficiency" and "cost".
RNGD adopts 5-nanometer tyrene process and is equipped with Furiosa's self-developed Tensor Contraction Processor architecture, which is designed to optimize tasks and has a computing performance of up to 512 TFLOPS with FP8 accuracy. The memory is configured with a dual-layer HBM3 stack, with a total capacity of 48GB, a bandwidth of 1.5TB/s, and an overall thermal design power consumption (TDP) of only 180 watts, far below the Nvidia H100 and up to 700 watts.
Although Nvidia's H100, B100 and other flagship GPUs are not against Nvidia's H100, B100, in terms of overall computing power and memory scale, RNGD has outstanding performance in "per-watt performance". LG test pointed out that RNGD generates up to 3.75 times more tokens than the A100 under the same power consumption conditions, showing the energy consumption advantages of far-reaching traditional GPU architecture.
In addition, RNGD avoids relying on advanced packaging and high-speed NVLink, and uses PCIe communication and combines software layer optimization to reduce overall deployment and integration costs, especially suitable for enterprise application scenarios that focus on recommendations. Compared with GPUs with training tasks as the core, the accelerator architecture is gradually developing towards "task-oriented" and "energy-efficiency-oriented".
It is worth mentioning that FuriosaAI once attracted Meta to buy. This also points out that the AI chip market will no longer be led by a single "universal" product, but will be a task-oriented low-power custom-made chip (ASIC). For enterprises such as LG that focus on visual energy and operating costs, RNGD provides a practical solution to effectively implement AI services by experiencing high-energy-consuming GPUs.
LG AI Research taps FuriosaAI to achieve 2.25x better LLM inference performance vs. GPUs How AI chip upstart FuriosaAI won over LG with its power-sipping design Extended reading: Korea's new creation FuriosaAI cooperates with LG to launch high-performance AI chip RNGD