Forbes - AI Chip Startup Cerebras Systems Looks To Challenge Nvidia's Dominance
Cerebrus, led by CEO Andrew Feldman, is an AI chip company that differentiates itself from NVIDIA by designing chips specifically for AI applications. Unlike NVIDIA's GPU heritage, Cerebrus started from scratch to create chips that are significantly faster and more energy-efficient. Their chips are large, the size of a dinner plate, which allows them to keep data on the chip, reducing the need for data movement and thus saving time and energy. This design choice addresses the fundamental AI computing challenge of frequent data movement. Cerebrus claims their chips can be up to 70 times faster than competitors like Meta's models, providing significant time savings for users. The company has successfully attracted major clients by offering solutions that are faster, cheaper, and require less effort than NVIDIA's offerings. Additionally, Cerebrus is involved in the open-source AI community, supporting models like Deep Seek, which have shown to be more efficient and capable of reasoning, thus broadening the accessibility and application of AI technologies.
Key Points:
- Cerebrus chips are designed specifically for AI, making them faster and more energy-efficient than NVIDIA's GPUs.
- Their large chip design reduces data movement, saving time and energy, crucial for AI computations.
- Cerebrus claims their chips can be up to 70 times faster than some competitors, significantly reducing response times.
- The company has attracted major clients by offering cost-effective and efficient AI solutions.
- Cerebrus supports open-source AI models, enhancing accessibility and innovation in AI technology.
Details:
1. 🎙️ Introduction & Guest Overview
- Katherine Schwab, Assistant Managing Editor at Forbes, focuses on technology.
- Andrew Feldman, co-founder and CEO of Cerebrus, an AI chip company, is the guest.
- The segment sets the stage for a discussion on technology and AI advancements, specifically around AI chip development.
2. 💡 Cerebrus vs. NVIDIA: A New Approach
- Cerebrus, an eight-year-old company, is planning an IPO later this year, signaling its readiness for market expansion and increased visibility.
- Cerebrus offers an innovative approach in chip technology that challenges the dominance of NVIDIA, the current leader in AI and chip manufacturing.
- The company differentiates itself by focusing on specialized chip designs that potentially offer more efficient AI processing capabilities compared to NVIDIA's general-purpose GPUs.
- Cerebrus' strategy includes targeting niche markets that require customized processing solutions, which may lead to significant shifts in how AI workloads are managed.
- This approach not only represents a technical innovation but also a strategic market positioning that could disrupt existing market dynamics dominated by NVIDIA.
- Cerebrus' innovations could lead to a new wave of chip technologies that prioritize efficiency and specialization over broad applicability.
3. ⚡ Speed & Efficiency of Cerebrus Chips
- Cerebrus chips are designed from scratch specifically for AI, unlike Nvidia's GPUs which have a different heritage, giving them a unique advantage.
- The design results in Cerebrus chips being significantly faster, with some tests showing performance improvements of up to 10 times compared to traditional GPUs.
- Cerebrus chips integrate chip, system, and software optimized for AI tasks, further enhancing their speed and efficiency.
- In real-world applications, Cerebrus chips achieve faster processing times, reducing AI computation from hours to minutes.
- The chips also offer improved power efficiency, consuming 30% less power than competing AI chips.
4. 🔍 The Science Behind Massive Chips
- Deep Seek, a leading AI model, demonstrates significant speed advantages, being up to 57 times faster than OpenAI's top models.
- In a direct comparison with Meta's LLaMA model, which boasts 70 billion parameters, Deep Seek outperforms by being 70 times faster.
- This speed disparity can drastically improve user experience, offering responses in as little as one second compared to 24 seconds with other models.
5. 🔄 Redefining Chip Design for AI
5.1. Challenges in AI Computation
5.2. Solutions Through Large Chip Design
6. 🏆 Competing with Industry Giants
6.1. Technical Innovations in AI Chip Development
6.2. Market Strategy and Competitive Positioning
7. 🌐 Deep Seek and Its Impact
- Deep Seek's recent model training used significantly fewer resources, impacting Nvidia's stock negatively.
- The introduction of Deep Seek has led to widespread discussions, even among those previously uninterested in AI.
- Deep Seek consists of two models: a giant parent model V3 and a distilled, smaller version for ease of use.
- The models were released with a permissive open-source license, facilitating widespread adoption.
- Several U.S. companies, including those at Perplexity and Together, began using these models on their equipment, showing high demand.
- Deep Seek's model is the first open-source model to allow reasoning, improving performance in math and coding tasks.
- The open-source nature of Deep Seek encourages innovation and customization in AI applications across various industries.
- Notable companies have quickly integrated Deep Seek into their operations, enhancing AI capabilities and reducing costs.
8. 🌍 Global Perspectives on AI Innovation
- The Deep seek 70b llama model, which integrates learnings from other models, is one of the top-performing open-source models, showcasing the potential of combining methodologies.
- The model's creation in China and its subsequent use by American companies highlights the global collaboration and innovation in AI, with open-source models being pivotal in this landscape.
- There is recognition that high-quality AI research and development are not confined to Silicon Valley, with significant contributions coming from universities and engineers in India, Europe, Dubai, Abu Dhabi, and China.
- Despite US regulations limiting computational resources, innovative techniques were developed and shared openly in the community, reinforcing the collaborative nature of AI advancement.
- Other regions, like Paris and Canada, are also producing noteworthy AI models, such as Misil and Coher, indicating a diverse and rich global AI ecosystem.
9. 📉 Market Responses and Misunderstandings
- Nvidia's stock decline is a result of a common misunderstanding in public markets: reducing prices and accelerating technology are often perceived as market shrinkers, while historically, these factors lead to market expansion. For instance, the PC and smartphone industries saw massive growth following similar trends.
- Emerging competitive AI models from China, such as the family of 100 models from Ali, challenge the belief that only a few companies can spearhead innovation. This illustrates the global potential in AI development, where innovation is not limited to a few major players.
- The global spread of AI capabilities, with advancements in China, Europe, India, and the Middle East, underscores that significant technological advancements can be achieved by teams of 100-200 skilled individuals at reasonable costs. This democratization of technology development suggests broader market impacts.
- Healthy markets thrive with multiple competitors rather than a few dominant players. This diversity fosters market expansion, as seen in historical cases like the automotive and telecommunications industries, which grew significantly with multiple active companies.
10. 🔄 Shifting Dynamics in AI Computing
- The traditional belief in AI is that more data and compute result in better AI models, but this requires significant financial investment.
- Creating cheaper AI models lowers the barrier to entry, making AI technology more accessible.
- Despite cost reductions, it's acknowledged that training these models is still expensive, but less so than previous models of similar size.
- Reasoning models use more compute during inference, leading to better accuracy but increased compute demands.
- There is a shift from compute dominance in training to increased compute use during inference.
- As AI model usage grows among users, the compute consumption rises significantly.
- The demand for faster response times from AI models is driving the need for more compute power.
- The AI ecosystem is rapidly advancing, which is beneficial for businesses involved in compute infrastructure.
11. 🇺🇸 US-China AI Race and National Security
- The Chinese government significantly invests in AI startups, refunding unsuccessful ventures to encourage venture capitalist participation, thereby bolstering its AI sector with robust state support.
- In contrast, the US AI sector relies more on venture capital, supplemented by government initiatives like the Chips Act, which supports the domestic chip ecosystem crucial for AI advancement.
- The Chips Act's $40 billion funding is dwarfed by private investments, such as Microsoft's $80 billion capex plan, highlighting the dominant role of private sector investment in the US.
- For the US to maintain its AI leadership and ensure national security, sustained governmental support is essential, focusing on keeping manufacturing and chip production within domestic borders.
12. 🛠️ Entrepreneurial Reflections & Future Plans
- The entrepreneur has been a CEO for 18 years and this is their fifth startup, highlighting their extensive experience in building companies.
- They emphasize a passion for tackling new, uncharted areas, likening themselves to those who seek 'dragons' as depicted on old maps, rather than staying in safe, known territories.
- The entrepreneur values creativity and insight in battling against large, established companies, indicating a preference for innovation and strategic challenges.
- They express a fear not of failure, but of being ordinary, suggesting a drive to pursue extraordinary achievements rather than settling for mediocrity.
- Their chip strategy had previously failed with major companies like IBM and Texas Instruments for 55 years, demonstrating their willingness to take on long-standing challenges.
- The entrepreneur succeeded where others had not, achieving something unprecedented in the computer industry, showing a commitment to groundbreaking work.
13. 👋 Closing Remarks & Gratitude
- The session concludes with expressions of gratitude towards Andrew for his time and insights.
- A note of thanks is extended to Forbes for hosting the discussion.
- Acknowledgment of future discussions implies ongoing collaboration or communication.
- Best wishes are offered for the success of Andrew's 'massive chips,' indicating a focus on or development of significant projects or technology.