AI semiconductor startup Furiosa AI is accelerating its advance into the global enterprise market by successfully mass-producing its second-generation AI-dedicated chip, 'RNGD (Renegade).'

Furiosa AI announced on the 28th that it has started receiving the first batch of 4,000 RNGDs from its foundry partner, TSMC, and will begin shipping card-type products in earnest this week.

The company plans to mass-produce a total of 20,000 units this year.

RNGD was first unveiled at the global semiconductor academic event 'Hot Chips 2024' held at Stanford University in the second half of 2024, and has now entered the mass production stage after undergoing performance verification and productization.

In particular, cases where a high-performance NPU equipped with HBM has gone beyond research and development and into actual mass production are rare even in the global semiconductor industry, and it is attracting attention as a new alternative in the AI infrastructure market that is highly dependent on GPUs.

The product comes in two forms. The 'RNGD PCIe card' is a drop-in AI accelerator that can be installed directly into existing servers thanks to its low-power design of 180W TDP.

The 'NXT RNGD Server' is a 4U rackmount server equipped with eight RNGD cards, and the total power consumption of the system is only 3kW.

It provides AI inference performance of up to 20 PFLOPS (INT8) per rack in a standard rack environment, securing both power and space efficiency.

Furiosa AI's strategy is to use this mass production as an opportunity to accelerate its expansion into the enterprise market.

Already, orders for RNGD adoption are pouring in from global companies, including affiliates of major domestic corporations, and cases of official adoption are rapidly increasing after verification in real-world environments.

The company has been accumulating practical results, such as the introduction of LG EXAONE and the demonstration of the OpenAI GPT-OSS model, while simultaneously stabilizing hardware and improving software stacks.

There is growing anticipation in the AI industry that RNGD could be a viable alternative to addressing data center infrastructure bottlenecks.

While GPU-centric infrastructures are increasing the power and cooling burden with power consumption exceeding 600W per chip, RNGD provides a 2.5x higher compute density per rack than GPUs while utilizing existing air-cooled infrastructure.

This means that it is possible to simultaneously address two challenges: reducing total cost of ownership (TCO) and expanding AI inference.

“The mass production of RNGD is a significant milestone for our leap forward in the global AI and semiconductor markets,” said Baek Jun-ho, CEO of Furiosa AI. “We will expand sales in the global enterprise market through AI infrastructure innovation based on power efficiency.”