3 Themes From Nvidia’s GPU Technology Conference (GTC)
This week we attended Nvidia’s s GPU Technology Conference (GTC) in San Jose. The conference grew to over 8,500 attendees, up 4X in five years. Long lines for the keynote zigzagged across the convention room floor, and the excitement was palpable. During the conference we noted three core themes: 1) automotive, 2) data center, and 3) crypto mining.
The themes suggest that next-generation hardware and ML applications will only grow in breadth. We continue to be excited by start-ups that accelerate AI/ML adoption at different layers of the stack from the edge to the data center.
- Automotive Remains a Large Market
Yesterday Nvidia traded off in the wake of news that it was suspending self-driving vehicle testing due to the fatal Uber crash in Arizona earlier this month. Uber and Toyota also said they would scale back on the road testing. During the GTC keynote Nvidia demonstrated an alternative, its new self-driving simulation. CEO Jensen Huang noted that while a fleet of 20 cars could only drive roughly 1 million miles in a year, the Constellation simulator could allow for billions of miles driven in a year without endangering pedestrians.
Nvidia believes its market opportunity expands as vehicle complexity moves beyond infotainment to ADAS and fully autonomous systems. Previously, the company forecast that its TAM will increase from $2 billion for Level 2/3 to $5 billion for Level 4/5 autonomy. Nvidia is well-positioned with Pegasus and Drive and announced 370 partners in the automotive space, up from 200 partners last year.
While the incident may present a temporary set-back, we believe there are interesting opportunities around the edge that benefit not just AV but also connected cars like secure V2V and V2X communication, in-vehicle security, and predictive maintenance.
2. Data Center Opportunities Exist Today
This year the company announced the world’s largest GPU, the Nvidia DGX-2, which weighs about 350 pounds and includes 16 individual Tesla V100 GPUs, each with 32 GB of system memory. The chipset is 10 times faster than the company’s DGX-1 that was announced last year. Nvidia’s NVSwitch tie together the GPUs that support two petaflops of computational power. DGX-2 targets deep learning applications.
Yesterday Nvidia increased its data center TAM assessment from $30 billion in 2020 to $50 billion in 2023 (HPC — $10 billion, Hyperscale and Consumer Internet — $20 billion, and Cloud Computing and Industries — $20 billion). Across these categories, Nvidia had stated that deep learning was a $11B TAM and inference was a $15B opportunity. The company also pointed out there are ~820,000 GPU developers and have been 8 million CUDA downloads.
The staggering explosion of interest in AI/ML catalyzed a proliferation of ML-tailored chips over the past few years. We’ve cataloged over 20 start-ups attempting to address the data center market. Despite general industry consolidation and possible mega-mergers of incumbent vendors, the largest semiconductor start-up exit within recent memory was Intel’s acquisition of Nervana for $350-$400 million. Other notable exits occurred in the connectivity space including Aquantia’s IPO valuation around ~$300 million and Cisco’s acquisition of Leaba for $320 million. With Cerebras’ latest post-money valuation at ~$860 million, we are excited to see a new wave of semiconductor companies address Nvidia’s ~$136 billion market cap.
3. Surprising Lack of Crypto Commentary
CEO Jensen’s talk didn’t announce or even allude to the Turing-series cryptocurrency mining GPUs. In the context of crypto, Jensen stated, “Gaming is growing, workstation is growing, AI hyperscale data center is growing, high-performance computing is growing. Quite frankly, I’d prefer that our GPUs were built to be used in those areas.”
Nvidia previously stated that crypto demands are a small percentage of its overall business but have resulted in supply problems for developers and gamers. Operators of distributed ledgers and cryptocurrency mining efforts favor the parallel processing found in GPUs because it helps meet the higher hash rates the market requires.
Ethereum mining using GPUs has become less popular because it currently isn’t as profitable and many miners have switched to ASICs for mining. According to RBC, assuming Ethereum Proof-of-Stake goes online in Q4, “miners would need to see a payback period of ~9 months to make their investment back before the algorithm changes. Overall, this would require a price of ~$900/ether.” At time of writing, the price was ~$457/ether suggesting that miners are unlikely to will add incremental equipment. All mining will not go to zero as numerous other tokens require memory hard hashing including Monero, Zcash, and Vertcoin.
While there could be some near-term choppiness in GPU purchasing surrounding the volatility in crypto, GTC highlights Nvidia’s expansive reach from autonomous vehicles to deep learning training. We are bullish about the future of ML infrastructure and are interested in speaking with start-ups tackling challenges in this space.