FuriosaAI just said “no thanks” to Meta’s hefty $800 million buyout offer. The South Korean chip maker is betting big on its RNGD chip instead – and for good reason. This beast pumps out 512 TFLOPS of processing power while sipping half the energy of Nvidia’s H100. With 48GB of HBM3 memory and air cooling, it’s a data center’s dream. Meta’s loss might be the AI industry’s gain, as this game-changing chip shows serious promise.
While Meta’s deep pockets have secured many tech acquisitions over the years, FuriosaAI isn’t taking the bait. The South Korean chip maker just turned down Meta’s whopping $800 million acquisition offer, choosing instead to forge ahead with its revolutionary RNGD chip. Guess Mark Zuckerberg can’t buy everything.
And who can blame them? The RNGD chip is a beast. Packing 512 TFLOPS of processing power at FP8 precision, it’s specifically designed for large language models and multimodal AI applications. Each chip features 20 RNGD accelerators that can be installed in a single server for increased processing power. But here’s the kicker – it does all this while consuming way less power than Nvidia’s energy-guzzling H100 GPU. We’re talking 150-200W versus 350W. Not too shabby.
FuriosaAI’s RNGD chip delivers 512 TFLOPS while using half the power of Nvidia’s H100. That’s game-changing efficiency.
Built on TSMC’s 5nm process, the RNGD chip isn’t just another pretty silicon face. It’s got 48GB of HBM3 memory, achieves 1.5 TB/s memory bandwidth, and features 256MB of SRAM for lightning-fast data access. The secret sauce? A fancy thing called the Tensor Contraction Processor that handles AI computations more efficiently than traditional methods.
The timing couldn’t be better. With data centers struggling to manage their power consumption and heat output, FuriosaAI’s air-cooled solution hits a sweet spot. The chip can process up to 3,000 tokens per second for 10-billion parameter models, all while keeping its cool. Literally. Their comprehensive software toolkit includes everything from model compression to debugging capabilities.
FuriosaAI’s already sampling the chip, with broader availability planned for early 2025. They’re clearly betting they can shake up the GPU-dominated AI hardware market on their own terms. And with specs like these, they might just pull it off.
The chip’s focus on sustainable AI compute, combined with its PCIe card format for easy deployment, makes it an attractive option for companies looking to green up their AI operations.
Meta’s loss might just be the tech world’s gain. Sometimes the most interesting innovations come from companies brave enough to say “no” to Silicon Valley’s biggest players.
