South Korean AI startup FuriosaAI has secured a significant deal for its RNGD (pronounced “Renegade”) energy efficient AI accelerator with an unlikely partner: fellow South Koreans LG Electronics, known for its consumer products.
Furiosa says its chips have successfully passed LG AI Research’s performance tests with its EXAONE models and LG will offer a product called RNGD Server to enterprise customers deploying LLMs. These include the diverse spectrum of LG businesses across electronics, finance, telecommunications, and biotechnology.
RNGD offers a Tensor Contraction Processor (TCP) chip architecture to deliver up to 512 TFLOPS of FP8 performance with a Thermal Design Power (TDP) of just 180W, making it far more power efficient than a GPU, which is its main selling point.
“After extensively testing a wide range of options, we found RNGD to be a highly effective solution for deploying EXAONE models. RNGD provides a compelling combination of benefits: excellent real-world performance, a dramatic reduction in our total cost of ownership, and a surprisingly straightforward integration,” Kijeong Jeon, lead, product unit, LG AI Research said in a statement. “For a project of this scale and ambition, the entire process was quite impressive.”
RNGD Server consists of up to eight RNGD accelerators in a single, air-cooled 4U chassis. FuriosaAI designed the server while LG acts as a solutions provider, delivering RNGD-powered servers to end customers.
LG is more typically associated with refrigerators and TV monitors, but there’s more to it than that. LG AI Research is the artificial intelligence think tank and research institute of LG Group that aims to develop cutting-edge AI technologies and solve complex challenges both within LG’s diverse business affiliates and in broader industry contexts. Its headquarters are in South Korea, but it has an office in Michigan.
The RNGD Server is now available to enterprise customers looking to develop large language models (LLMs) like LG’s EXAONE LLM which is being deployed across sectors like electronics, finance, and telecommunications.
In LG AI Research’s testing, RNGD achieved 2.25 times better performance per watt for LLMs compared to an unnamed GPU-based package. Thanks to greater compute density, a RNGD-powered rack can generate 3.75 times more tokens for EXAONE models compared to a GPU rack operating within the same power constraints.
There are dozens of AI chip startups out there, but FuriosaAI has gotten some attractive attention. Meta, the parent company of Facebook, earlier this year offered buy out the company for $800 million, but FuriosaAI decided to pass on the deal.
Still, there are several noteworthy accelerators that are specializing in AI inference, including Cerebras, Graphcore, Groq, and SambaNova, and others. So, it’s a very competitive market, said Addison Snell, CEO of Intersect360 Research.
“It’s hard to imagine any of them carrying enough weight to be a real threat to Nvidia, but the market is vibrant enough for any of them to rack up some significant wins,” he said.
Source:: Network World