Acquired NVIDIA Part II: The Machine Learning Company (2006-2022)
Category: Acquired Podcast (Deep Dives) · Duration: 135 min · ▶ Watch
Speakers: Ben Gilbert and David Rosenthal
Segments (27)
- 00:00 · Introduction & The Scale of Compute
- The hosts introduce the episode and discuss the unfathomable scale of computations required for modern AI.
- 07:00 · Nvidia’s Early Strategy & Execution
- A recap of Nvidia’s early days, focusing on their aggressive 6-month product cycles and decision to write their own drivers.
- 15:00 · The Stanford Researcher & Early GPU Compute
- The story of a researcher using GPUs for quantum chemistry, highlighting the early potential of GPUs for general compute.
- 21:00 · The Creation of CUDA
- Jensen Huang’s massive bet on creating CUDA, a software layer to make GPUs programmable for general tasks.
- 24:00 · The 2008 Crash & Staying the Course
- Nvidia’s stock plummets 80% during the financial crisis, but the company remains committed to its expensive CUDA strategy.
- 36:00 · The Tegra Mobile Experiment
- Nvidia’s attempt to enter the mobile processor market with Tegra, which ultimately found success in the Nintendo Switch.
- 43:00 · ImageNet and the AI Big Bang
- The creation of the ImageNet dataset by Fei-Fei Li, setting the stage for the deep learning revolution.
- 46:00 · AlexNet Changes Everything
- Alex Krizhevsky and team use Nvidia GPUs to train a neural network that shatters previous ImageNet records.
- 45:50 · The AlexNet Breakthrough
- The hosts discuss how the AlexNet team used deep learning and GPUs to achieve a massive breakthrough in AI image recognition.
- 48:21 · CUDA and the Search for a Market
- Nvidia’s development of CUDA and cuDNN provided the essential software layer that made GPUs accessible for AI researchers.
- 51:41 · The Trillion Dollar AI Market
- The realization that AI could revolutionize digital advertising and content aggregation created a massive new market for Nvidia’s hardware.
- 54:41 · Nvidia’s Stock Journey and Crypto Boom/Bust
- A look back at Nvidia’s stock price history, including the massive run-up and subsequent crash driven by cryptocurrency mining demand.
- 57:21 · Parallel Computing: Graphics, AI, and Crypto
- The core realization that graphics, neural networks, and crypto mining all rely on ‘embarrassingly parallel’ matrix math perfectly suited for GPUs.
- 01:10:21 · The Rise of the Data Center Segment
- Nvidia’s Data Center revenue exploded, matching and eventually surpassing its core gaming business due to enterprise AI demand.
- 01:20:21 · The Mellanox Acquisition and Data Center Scale
- Nvidia acquired Mellanox to control the high-speed networking required to connect thousands of GPUs, making the data center the new unit of compute.
- 01:24:21 · The Failed Arm Acquisition and Grace CPU
- Despite failing to acquire Arm, Nvidia pushed forward with developing its own Arm-based data center CPU, Grace.
- 01:34:21 · Nvidia’s Financials and Future Outlook
- An analysis of Nvidia’s high valuation, massive free cash flow, and strategy to sell enterprise software solutions.
- 91:40 · The Omniverse Vision
- The hosts discuss Nvidia’s ambitious goal to simulate the physical world using their hardware and software.
- 94:55 · Vanta Sponsorship
- An ad read for Vanta, a compliance and security platform.
- 97:40 · Recap of Nvidia’s Early Days
- A brief review of Nvidia’s near-death experiences and early strategies in the graphics card market.
- 100:30 · The Birth of CUDA
- Discussion on Nvidia’s massive bet to make GPUs programmable for general-purpose computing.
- 106:00 · The Cost of Innovation
- How building CUDA added significant costs to chips without immediate benefit to their core gaming customers.
- 111:00 · Market Crash and Skepticism
- Nvidia faces a massive stock drop amid the 2008 financial crisis and Wall Street’s doubt about CUDA’s ROI.
- 115:00 · The Mobile Misadventure
- Nvidia’s failed attempt to dominate the smartphone market with their Tegra chips.
- 119:15 · Hyperscaler Capital Expenditure
- Analyzing the massive investments made by tech giants in data center infrastructure.
- 125:00 · The AI Big Bang
- How the AlexNet breakthrough proved that GPUs were the perfect hardware for deep learning.
- 130:35 · Carve Outs
- The hosts share their personal recommendations for books, cameras, and experiences.
Specific Prices (12)
| Timestamp | Item | Value | Context |
|---|---|---|---|
| 13:45 | Nvidia Market Cap | $20 billion | Nvidia’s peak market capitalization in mid-2007 before the financial crisis. |
| 23:45 | ATI Acquisition | $6-7 billion | The estimated price AMD paid to acquire ATI, Nvidia’s main graphics competitor. |
| 25:00 | Nvidia Stock Drop | 80% decline | The massive drop in Nvidia’s stock price during the 2008 financial crisis. |
| 54:41 | Nvidia Stock (2012-2015) | ~$5 | The price of Nvidia stock before the market realized its potential in AI. |
| 54:55 | Nvidia Stock (Current at recording) | ~$220 | The stock price at the time the podcast was recorded. |
| 55:18 | Nvidia Stock (2018 Peak) | $65 | The peak stock price during the 2017/2018 cryptocurrency mining boom. |
| 55:33 | Nvidia Stock (2019 Trough) | $34 | The stock price after the crypto crash. |
| 01:06:43 | RTX 3090 GPU | ~$2,000 | The price of a high-end consumer graphics card. |
| 01:07:06 | H100 GPU | $20,000 - $30,000 | The estimated price of a single enterprise data center GPU. |
| 01:10:00 | Mellanox Acquisition | ~$7 Billion | The price Nvidia paid to acquire networking company Mellanox in 2020. |
| 111:30 | AMD acquisition of ATI | ~$6-7 billion | The price AMD paid to acquire Nvidia’s main graphics competitor, ATI. |
| 112:30 | Nvidia Market Cap | Dropped from ~$20B to ~$4B | Nvidia’s valuation plummeted 80% during the 2008 financial crisis and due to market skepticism. |
Memory Facts (4)
- [19:50] Employees dedicated to the CUDA platform
- 1,100 employees
- [30:00] Number of cores on a modern consumer GPU
- Over 10,000 cores
- [46:41] Deep learning algorithms require massive amounts of compute.
- Described metaphorically as needing compute on the order of ‘grains of sand on earth’.
- [94:10] Training a single speech recognition model takes more math operations than grains of sand on Earth.
- Math operations vs. grains of sand
Bottleneck Claims (3)
- [16:30] Programming languages were the bottleneck for GPU compute.
- Evidence: Researchers had to translate complex math into graphical shader languages (CG) just to use the GPU hardware.
- [46:41] Deep learning was bottlenecked by computational power before GPUs were utilized.
- Evidence: The algorithms existed for decades but were impractical to run on traditional CPU architectures.
- [01:06:21] The current bottleneck in computing is moving workloads off the CPU.
- Evidence: The industry shift towards specialized accelerators (GPUs, DPUs) to handle tasks CPUs are too slow for.
Predictions (4)
- [22:30, Long-term (played out over 10+ years)] If Nvidia builds a general-purpose compute platform (CUDA), developers will eventually find uses for it.
- [01:00:21, Future/Ongoing] There will be more massive markets discovered that rely on parallel matrix multiplication.
- [01:33:21, Long-term] Nvidia is targeting a $1 Trillion total addressable market.
- [93:30, Long-term future] Nvidia’s Omniverse will allow for the simulation of the entire physical world.
Key Technologies (13)
- Programmable Shaders: Allows developers to write custom code to dictate how pixels and vertices are rendered, moving away from fixed-function graphics.
- CUDA (Compute Unified Device Architecture): A parallel computing platform and API model that allows software developers to use a CUDA-enabled GPU for general purpose processing.
- Tegra: Nvidia’s system-on-a-chip (SoC) series designed for mobile devices, integrating an ARM architecture processor and Nvidia GPU.
- Convolutional Neural Networks (CNNs): A class of deep neural networks, most commonly applied to analyzing visual imagery, which proved highly effective when run on GPUs.
- Deep Learning / Neural Networks: A branch of AI that uses multi-layered algorithms to learn from large amounts of data.
- CUDA: Nvidia’s parallel computing platform and programming model that allows developers to use GPUs for general-purpose processing.
- cuDNN: A GPU-accelerated library of primitives for deep neural networks, built on top of CUDA.
- Transformer Models: Advanced AI architectures used for tasks like natural language processing and image generation.
- NVLink / Infinity Fabric: High-speed interconnect technologies used to link multiple chips together with low latency.
- DPU (Data Processing Unit): A specialized processor designed to handle data center networking and data movement tasks.
- CUDA (Compute Unified Device Architecture): A parallel computing platform and programming model that allows developers to use GPUs for general-purpose processing.
- Tegra: Nvidia’s system-on-a-chip (SoC) series developed primarily for mobile devices.
- Deep Learning: A subset of machine learning based on artificial neural networks that requires massive parallel compute power.
Companies Mentioned (19)
Intel · Microsoft · AMD · Nintendo · Google · Facebook (Meta) · Baidu · Cerebras · Tesla · Mellanox · Arm · TSMC · Vanta · Nvidia · ATI · Apple · Qualcomm · Microsoft, Amazon, Meta, Alphabet · Softbank
Notable Quotes (7)
Thank you, I can get my life’s work done in my lifetime. — Jensen Huang (paraphrased) @ 15:30
If you don’t build it, they can’t come. — Jensen Huang (paraphrased) @ 22:30
Embarrassingly parallel. — Ben Gilbert @ 30:45
We cannot overstate the importance of this moment… This was the Big Bang moment for artificial intelligence. — Host @ 48:00
If we were a hedge fund, we’d put all our money into Nvidia. — Host (quoting Marc Andreessen) @ 54:08
You say solutions, I hear gross margin. — Host @ 01:21:08
If you don’t build it, they can’t come. — Host (paraphrasing Jensen Huang) @ 108:20
Key Topics
The historical evolution of Nvidia from a gaming company to an AI infrastructure giant. · The strategic importance and massive financial risk of developing the CUDA platform. · The serendipitous intersection of GPU architecture and the computational needs of deep learning. · The impact of the AlexNet breakthrough on the tech industry. · Artificial Intelligence · GPU Computing · Nvidia Corporate History · Data Center Infrastructure · Semiconductor Industry · Nvidia's strategic long-term bets · The development and financial burden of CUDA · Nvidia's failure in the mobile chip market · The intersection of GPUs and the rise of Deep Learning · The massive scale of hyperscaler capital expenditure
Takeaways
- Nvidia’s current dominance in AI was not an overnight success, but the result of a decade-long, highly risky investment in the CUDA software ecosystem.
- Hardware alone is insufficient; Nvidia’s moat is built on the software layer (CUDA) that makes their hardware accessible to researchers and developers.
- The AI revolution was catalyzed when researchers realized that the ‘embarrassingly parallel’ nature of graphics processing perfectly matched the math required for neural networks.
- Nvidia’s GPUs, originally designed for graphics, proved to be the perfect hardware for the parallel processing required by deep learning.
- The creation of CUDA and cuDNN created a massive software moat that locked developers into Nvidia’s ecosystem.
- Nvidia successfully transitioned from a gaming-focused company to a data center and AI powerhouse.
- The acquisition of Mellanox allowed Nvidia to scale their architecture from individual chips to entire data centers.
- Nvidia’s current dominance in AI is the result of a decade-long, highly criticized investment in making GPUs programmable via CUDA.
- Strategic failures, such as Nvidia’s attempt to enter the mobile market, can provide valuable lessons and technology that pivot into future successes.
- Building a robust software ecosystem around hardware creates a significantly deeper competitive moat than hardware alone.