As its revenue shrinks, Nvidia bets the farm on AI

Artificial intelligence is the “single most powerful force of our time,” says CEO Jensen Huang

by Max Smolaks 16 August 2019

American chipmaker Nvidia has been talking up the potential of artificial intelligence, as it faces increased competition from AMD in its traditional video gaming domain, and the collapse of the cryptocurrency mining rush that propped it up in 2018.

The company has reported $2.579 billion in revenue for the second quarter of its financial 2020 – down 17 percent year-on-year, but still ahead of analyst expectations. Net income stood at $552m, down 50 percent y-o-y.

CEO Jen-Hsun ‘Jensen’ Huang underscored the importance of the AI sector for the future of Nvidia during an earnings call, identifying the technology as the “single most powerful force of our time.”

Do it like Walmart

“We are building a broad base of customers across multiple industries as they adopt Nvidia’s platforms to harness the power of AI. Public sector, higher education and financial services were among the key verticals driving growth this quarter,” he said.

“In addition, we won Lighthouse account deals in important industries that are on the cusp of being transformed by AI. For example, in retail, Walmart is using Nvidia GPUs to run some of its product demand forecasting models, slashing the time to do so in just 4 hours from several weeks on CPUs.”

Nvidia made its fortunes by selling purpose-built silicon to PC gamers, but the rising popularity of AI workloads in HPC environments and corporate data centers has motivated it to shift its focus to enterprise computing. The company launched its own AI server range under the DGX brand in 2016, and recently built a supercomputer with 1,472 GPUs that debuted as the 22nd fastest machine in the world – according to the latest Top500 supercomputer list.

Earlier this week, this machine – dubbed DGX SuperPOD – set a world record for advanced natural language processing across both BERT (Bidirectional Encoder Representations from Transformer) and GPT-2 (Generative Pretrained Transformer 2) models.

“This can enable more human comprehension in real-time by chat bots, intelligent personal assistants and search engines. We are working with Microsoft as an early adopter of these advances,” Huang said.

During the earnings call, he noted there were currently more than 4,000 AI start-ups around the world.

“The next phase of AI is what we call vertical industry enterprise AI,” Huang said. “And this is where companies are using it not just to accelerate the business process internally, but they’re using AI to create new products and services.

“They could be new IoT-based medical instruments to monitor your health. It could be something for financial services, for forecasting or for fraud detection. It could be some kind of device that delivers pizza to you. And with the combination of IoT and artificial intelligence, for the very first time, you actually have the software capabilities to make use of all of these sensors that you’re putting all over the world.

“And so that phase of growth of AI is the phase that we’re about to enter into.”

Chatbots can’t wait

Huang said Nvidia was particularly interested in chatbots, since they require a ton of processing capacity. “The thing that’s really exciting for us is beyond recommendations, beyond image enhancement, the area where we believe the most important application for AI is likely conversational AI,” he said.

“Photo tagging and photo enhancement is something that you could put offline and you could do that while you have excess capacity, avoiding the busiest time of the day. You can’t do that with language and conversational AI. You better respond to the person in real-time, and so the performance that’s required is significant.

“But more importantly, the number of models necessary for conversational AI – from speech recognition to language understanding to recommendation systems to text-to-speech to wave synthesis – these 5, 6, 7 models have to be processed in real-time, in a series so that you can have a reasonable conversation with the AI agent.

“My expectation is that this is going to continue to be a big growth opportunity for us.”

Huang said Nvidia continued to advance its software, built around the CUDA programming model, and managed to increase AI performance of its Volta GPUs by 80 percent in 12 months. This meant that customers continued to see major gains in speed, long after they paid for the product – because of their complex relationship between hardware and software, GPUs are pretty much the only class of device capable of doing this.

Revenue breakdown

In terms of particular business segments, Nvidia’s gaming revenue – still at the core of its empire – was $1.31 billion, down 27 percent from a year ago. “The year-on-year decrease reflects a decline in shipments of gaming desktop GPUs and SOC modules for gaming platforms, partially offset by growth in gaming notebook GPUs,” Colette Kress, Nvidia’s CFO, explained [PDF].

Nvidia’s data center segment didn’t do nearly as badly, with quarterly revenue of $655 million, down 14 percent from a year ago, and up 3 percent sequentially. “The year-on-year decline reflects lower hyperscale revenue. The sequential increase was due to enterprise revenue growth driven by expanding AI workloads,” Kress said.

Professional visualization product revenue was $291 million, up 4 percent, automotive revenue hit a record $209 million, up 30 percent, and the Tegra chip family, designed for mobile devices and powering Nintendo’s Switch console, brought in $475 million, up 2 percent year-on-year.

OEM and “other” revenue totaled $111 million, down 4 percent from a year ago and up 12 percent sequentially, with the sequential increase “primarily due to growth in shipments of embedded edge AI products,” according to Kress.

Nvidia is about to close the acquisition of Israeli chipmaker Mellanox, famed for its data center networking kit and the InfiniBand interconnect, used to link together individual nodes in high performance computing. The GPU giant is spending a whopping $6.9bn on the purchase – which is once again aimed at boosting its enterprise credentials, rather than entice gaming enthusiasts.