How Did Nervana Build the First AI Chip Startup?

TL;DR

Nervana Systems emerged after Naveen Rao concluded that artificial neural networks needed a computing substrate designed specifically for their workloads. His combined experience in chip design, software, machine learning, and computational neuroscience helped the startup choose a practical architecture, attract support as attitudes toward hardware evolved, and eventually become part of Intel through an acquisition valued at more than $400 million.

Transcript

welcome to fyi the four-year innovation podcast this show offers an intellectual discussion on technologically enabled disruption because investing in innovation starts with understanding it to learn more visit arc dash dashinvest.com arc invest is a registered investment advisor focused on investing in disruptive innovation this podcast is for inf... Read More

Key Insights

• Naveen Rao’s technical foundation is a combination of electrical engineering, computer science, chip architecture, software development, machine learning, and computational neuroscience. That range allowed him to connect research about biological intelligence with the practical requirements of designing specialized computing systems.
• The efficiency gap between computers and the human brain motivated Rao’s long-term interest in intelligent machines. He observed that a computer operating at 20 watts could not approach the brain’s capabilities, making energy-efficient intelligence appear highly difficult during his undergraduate years in the 1990s.
• Computational neuroscience studies how the brain transforms information into behavior using mathematical and computational models. Rao worked with observations from behaving animals and machine-learning methods to investigate processes such as converting visual signals into motor plans and motor programs.
• Deep learning became practical through the convergence of established techniques, growing data availability, and greater computational power. Rao saw that backpropagation and related methods were beginning to reproduce some aspects of intelligence, creating a need for computing hardware better suited to neural-network workloads.
• Neuromorphic computing represents information through neuron-like pulses called action potentials instead of the two’s-complement mathematics used by standard logic systems. Qualcomm investigated whether this biologically inspired representation could produce vastly more efficient, intelligent computing systems that consumed very little power.
• Nervana’s architectural direction was driven by engineering pragmatism rather than strict biological imitation. Although Rao had studied neuroscience and neuromorphic machines, he believed conventional artificial neural networks offered a more practical and scalable foundation than the spiking-neural-network direction Qualcomm preferred.
• Nervana originated after Qualcomm declined Rao’s internal proposal to build hardware for artificial neural networks. Rao and colleague Amir Khosrowshahi discussed forming a company, sought encouragement from Berkeley professor Bruno Olshausen during a Christmas-break lunch, and began the startup process in 2014.
• NVIDIA’s use of GPUs for deep learning helped create an industry of specialized AI processors. The discussion says more than 50 companies appeared over roughly five years to accelerate AI in data centers, edge systems, and cars, while Nervana preceded later startups such as Cerebras, Groq, and Graphcore.

Questions & Answers

Q: How did Naveen Rao’s background lead to Nervana Systems?

Naveen Rao combined undergraduate training in electrical engineering and computer science with work as a computer architect and experience building ASICs and software stacks. He later earned a PhD in computational neuroscience and researched neuromorphic machines at Qualcomm. When deep learning gained momentum, that combined background helped him recognize the need for a computing substrate designed specifically for artificial neural networks.

Q: Why did Naveen Rao study computational neuroscience?

Rao wanted to understand whether concepts from biological brains could fundamentally change what computers can do. After designing computers and working at several startups, he left his job in 2007 to pursue a PhD in computational neuroscience. His research used data from behaving animals and machine learning to characterize how the brain converts inputs, including visual signals, into motor plans and programs.

Q: What was Qualcomm researching in neuromorphic computing?

Qualcomm was investigating whether neuron-like communication could enable extremely efficient intelligent systems. Biological neurons use action potentials, or pulses, to communicate information through rates and relative timing. This differs from the two’s-complement mathematics of conventional logic. The research tested whether representing computation and numbers in this biologically inspired way could support low-power intelligence and help Qualcomm move ahead of competitors.

Q: Why did Rao choose artificial neural networks over spiking neural networks?

Rao chose artificial neural networks because he considered them more practical and scalable, despite his background in neuroscience and neuromorphic research. At Qualcomm, he proposed building hardware for conventional neural networks while the company preferred to continue pursuing spiking approaches. He viewed himself primarily as an engineer who valued scientific research but ultimately wanted to build systems that could work at scale.

Q: How was Nervana Systems founded in 2014?

Nervana began after Qualcomm declined Rao’s internal proposal to develop hardware for artificial neural networks. Rao discussed the opportunity with Qualcomm colleague Amir Khosrowshahi, who became enthusiastic about the idea. They then met Khosrowshahi’s PhD adviser, Berkeley professor Bruno Olshausen, during the Christmas break. His strong encouragement helped convince them that the timing was right, and they started the company-building process in 2014.

Q: Why did deep learning create a market for specialized AI chips?

Deep learning began achieving stronger results when existing techniques such as backpropagation converged with larger quantities of data and increased computational power. NVIDIA’s use of programmable GPUs demonstrated that these workloads could be accelerated effectively. Rao concluded that conventional computers were not designed for this type of problem, so a more suitable computing substrate could improve how artificial neural networks were executed.

Q: What happened after Intel acquired Nervana Systems?

Intel acquired Nervana for more than $400 million, and Rao subsequently led Intel’s broader AI group. According to the description, the company retained some Nervana developments while also pursuing different directions. The experience revealed that joining a much larger semiconductor organization did not eliminate personnel and resource constraints, and Rao’s responsibilities changed as Nervana became integrated into Intel’s operations.

Q: How did Nervana fit into the growth of the AI chip industry?

Nervana is presented as the first specialized AI chip startup created after NVIDIA helped launch the modern deep-learning movement through GPUs. Over roughly five years, more than 50 companies appeared to accelerate artificial intelligence in data centers, edge devices, and cars. Nervana preceded companies including Cerebras, Groq, and Graphcore, which collectively attracted more than $2 billion in venture funding.

Summary & Key Takeaways

• Naveen Rao’s path to Nervana combined electrical engineering, computer science, semiconductor design, and computational neuroscience. After building ASICs and software stacks for video, content delivery, compression, and wireless networking, he pursued a PhD to investigate whether principles from biological brains could inform more intelligent and energy-efficient synthetic computing systems.

• At Qualcomm, Rao researched neuromorphic computing based on spiking neurons and biologically inspired signal representations. He eventually favored conventional artificial neural networks because he considered them more practical and scalable. When Qualcomm declined his proposal to build hardware for that approach, he treated the decision as a signal to establish Nervana Systems.

• Nervana entered a market created by the success of deep learning on NVIDIA GPUs and became the first of many specialized AI chip startups described in the discussion. Intel ultimately acquired Nervana for more than $400 million, after which Rao led Intel’s AI group and experienced the constraints of innovation inside a large semiconductor company.