AI chip developer FuriosaAI reportedly raising up to $500M – News

FuriosaAI Inc., a Seoul-based developer of artificial intelligence chips, is reportedly in talks to raise a new round of funding.

Sources told Bloomberg today that the startup is seeking $300 million to $500 million. It has reportedly hired Morgan Stanley and Mirae Asset Securities, Korea’s largest investment bank, to help it close the round. The proposed investment is described as a Series D raise.

The report didn’t specify the valuation that FuriosaAI is expected to receive. The company was worth $735 million as of July, when it raised a $125 million round from a group of institutional investors. A few months earlier, it reportedly turned down a $800 million acquisition offer from Meta Platforms Inc.

There are often multiple ways to go about a mathematical operation. For example, adding two and two produces the same result as multiplying two by two. A similar principle applies to matrix multiplications, the calculations that AI models use to process data.

A matrix is a mathematical structure that comprises a collection of numbers organized into rows and columns. A matrix multiplication combines two matrices into one through a process that involves multiplication, but also comprises other operations. Server-grade graphics processing units have numerous cores optimized specifically to perform matrix multiplications.

FuriosaAI has taken a different approach with its flagship RNGD chip (pictured). Instead of designing the processor with matrix multiplications in mind, the company optimized it for a mathematical operation called tensor contraction. FuriosaAI says that tensor contractions can produce the same results as a matrix multiplication but more efficiently, which translates into faster AI performance.

The company’s chip architecture speeds up neural networks in two main ways. First, it improves their parallelism, which means that more calculations can be carried out side-by-side instead of one after another. The technology also increases data reuse. Reusing a piece of data reduces the number of times it has to be retrieved from memory or generated from scratch, which avoids the associated delays.

The RNGD can perform up to 512 trillion calculations per second on data stored in the FP8 format. It ships in a PCIe card with a thermal footprint of 180 watts, which means that it doesn’t require water cooling. FuriosaAI also offers a server called the NXT RNGD that includes eight RNGD accelerators and provides up to 4 petaflops of performance.

The funding round that the company is raising could reportedly be its last before going public. FuriosaAI is expected to list its shares in 2027.

Photo: FuriosaAI