South Korean chip startup FuriosaAI has launched a brand new server product to deal with the corporate’s AI inference semiconductors.
The NXT RNGD (pronounced “renegade”) server will home the corporate’s neural processing RNGD chips. The corporate mentioned its servers will supply greater effectivity and decrease price in comparison with GPU options for on-premises knowledge middle environments and personal cloud.
In an interview with DCN, FuriosaAI CEO June Paik mentioned the corporate needs to carve out a distinct segment in inference processing by pioneering a turnkey resolution with a mixture of open-source and in-house software program choices. He mentioned the corporate is concentrating on prospects within the banking, finance, training, and e-commerce sectors, in addition to different enterprise prospects.
“The market proper now could be dominated by only one participant,” Paik mentioned. “Nvidia is kind of dominating. There are a number of corporations like AMD and Intel attempting to catch up… and there are such a lot of enterprise corporations that want AI, we want a brand new approach of doing this. Our ambition is to be the following modern computing firm globally.”
FuriosaAI CEO June Paik. Picture: FuriosaAI.
Tackling a Big
Taking over Nvidia is a tall order. The Santa Clara {hardware} firm at present dominates in AI coaching and inference, although it’s dealing with extra competitors for the latter. Startups like FuriosaAI, xAI, and Cerebras are making inroads within the inference area with specialised {hardware} and {custom} silicon.
Nvidia can be dealing with inference competitors from massive tech gamers – Google is utilizing custom-built v4i TPU chips, Amazon has its in-house Inferentia chips. Intel, AMD, and Qualcomm are additionally aiming on the inference compute market share.
“There are such a lot of deployment situations for inference, and I’ve little question that FuriosaAI and its RNGD Server will discover a place out there,” Matthew Kimball, vp and principal analyst at Moor Insights & Technique mentioned in an e-mail.
Kimball mentioned he’s trying ahead to seeing extra benchmarking outcomes from the servers.
The corporate mentioned NXT RNGD will present 384 GB of HBM3 reminiscence working at 12 TB/s reminiscence bandwidth with a complete energy consumption of three kW, in comparison with 10 kW or extra for high-powered GPU servers. The unit holds eight RNGD playing cards delivering 4 petaflops of FP8 (or 4 petaflops of INT8) compute.
The corporate claims a typical 15 kW knowledge middle rack can maintain as much as 5 NXT RNGD Servers, in comparison with one Nvidia DGX server.
Difficult CUDA Dominance
Regardless of the decrease price, there’s nonetheless an enormous hurdle for FuriosaAI and different potential Nvidia rivals. Nvidia’s CUDA software program has develop into ubiquitous, giving builders a dependable and well-supported resolution.
Paik mentioned prospects might have develop into depending on CUDA for coaching workloads, however on the inference aspect, “there’s much less dependency,” and he sees a gap for FuriosaAI. “The problem is for us to exchange this CUDA engine with our personal software program stack,” he mentioned. “We’re doing fairly good with our personal engine now.”
FuriosaAI claimed an enormous buyer win in July with LG AI Analysis, which can supply RNGD-powered servers to enterprise prospects. Earlier within the 12 months, FuriosaAI reportedly turned down an $800 million bid from Meta. Paik declined to touch upon the reported supply.
The corporate’s world prospects are at present sampling the NXT RNGD server, and the items will likely be obtainable for order in early 2026.
