Nvidia’s influence in the AI chip space is ever-growing, not the least thanks to the success of its GPU software CUDA.
Amid this climate, South Korean chip startup FuriosaAI is aiming to carve its mark by developing both hardware and software.
Unless chip companies have a software stack that can react to various black box models, it will be very difficult for them to secure customers, FuriosaAI CEO June Paik said in a conference hosted by TheElec on Tuesday.
Paik stresses the importance of software as there are design limitations in AI chips.
In terms of hardware, it takes a minimum of two years for an AI chip to be designed and go into production. Two years might as well be decades in the fast-paced chip industry. This means that by the time an AI chip optimized for a certain AI service is made, it may be outdated.
That is why FuriosaAI is just as focused on software and it complies a number of multiple matrix units, or MXU, into a larger MXU, Paik said.
This allows the company to react to various forms of data, the CEO said.
FuriosaAI was founded in 2017 and designs neural processing units, and chips designed specifically to process AI. Last year, it started production of its first AI chip Warboy; it is also developing a Gen 2 chip aimed at data centers called Renegade.
Nvidia had approximately 40,000 enterprise customers as of its latest fiscal quarter, Paik said, which means it can’t design chips customized for individual customers.
So other AI companies besides Nvidia are focused on offering more power efficiency in their chips for specific fields compared to Nvidia’s, the CEO said.
Most data centers used Nvidia A100 and H100 as there is little alternative in the market right now.
Google, IBM, Meta, Microsoft, and other big companies are developing their own chips optimized for their own services.