Jensen Huang stated during Wednesday's earnings call that NVIDIA's custom AI chip, following its $20 billion acquisition of Groq's technology, will remain a niche product for some time.
NVIDIA unveiled the new Groq language processing unit and a complete data center rack filled with these custom chips, named LPX, at its annual GTC conference in March.
"LPX is designed for low latency and high token rates, but it has lower throughput," Huang said. "The use cases for LPX are not broad."
The new Groq chip in the LPX is a typical example of an Application-Specific Integrated Circuit (ASIC). These chips consume less power and are programmed for more specific tasks.
The high-profile public listing of custom ASIC maker Cerebras Systems on Thursday serves as the clearest signal yet of the market's appetite for alternatives to NVIDIA's GPUs.
ASICs are also manufactured in-house by companies such as Alphabet, Amazon.com, Meta Platforms, Inc., and Microsoft, while a growing number of startups like SambaNova and D-Matrix are entering the field.
Comments