With Intel and AMD elbowing into its niche, Nvidia introduced a new line of graphics chips aimed at winning the title of fastest processors for running artificial intelligence programs.
Although the company gained prominence by selling some of the fastest graphics cards for playing video games, it has refocused, along with the market, to the more lucrative niche of running A.I. apps in corporate and cloud data centers. Last year, Nvidia sold almost $3 billion in data center chips, representing slightly more than one-quarter of its total sales.
But that market is growing increasingly competitive. Advanced Micro Devices has debuted new data center chips while Intel plans to start selling its own graphic chips sometime next year.
A.I. apps are all the rage as cloud computing centers run by Amazon and others increasingly are used to run programs recognizing speech or images, making diagnoses from medical tests, and seeking to learn from patterns in other big data sets.
The lead product in Nvidia’s new Ampere lineup, the A100, packs 54 billion transistors onto a single chip. The company says it can run A.I. programs that train using deep learning, like speech and image-recognition apps, at a speed of 312 trillion operations per second, or 20 times as fast as its three-year-old Volta line. That means not only could it save customers money by running their A.I. training programs more quickly, it could also enable new and even more complicated A.I. apps that weren’t practical on older, slower chips.
For more general high-performance computing tasks, the A100 can hit almost 20 trillion operations per second, about 2.5 times as fast as Nvidia’s older chips.
Nvidia CEO Jensen Huang said all of the major cloud service providers, including Amazon, Google, and Microsoft, had already agreed to install the new chips in their servers. “Everybody was instantaneously lined up,” he said on a call with reporters on Wednesday. Although data centers previously relied mostly on the same kinds of processors found in PCs, companies have found that using graphics chips accelerates many kinds of tasks such as A.I., he noted. “Data center acceleration is now conventional wisdom…It’s common sense,” Huang said.
Another key aspect of Nvidia’s new chips is a faster technology for connecting multiple graphics chips together and connecting to storage systems, a critical function in huge data centers made up of thousands of servers. Last month, Nvidia completed its $7 billion acquisition of Mellanox Technologies, which develops high-speed data connections for networking server and cloud computers.
Even before the arrival of the newest chips, Nvidia has been pounding its rivals in the stock market. Nvidia’s shares have gained 32% so far this year, beating AMD’s 14% rise and Intel’s 4% drop.
