Jensen Huang Took 20 Years to Complete GPGPU [Tech Talk]

What Did Jensen Huang Mean by His Quantum Computer Remark?
NVIDIA's AI Chip Also Took 20 Years to Develop
What Matters Is Not Just Technology, But Market Maturity

"I think many people would believe if we choose 20 years until a useful quantum computer emerges."

Jensen Huang, CEO of NVIDIA, recently gave this answer when asked about the commercialization timeline of quantum computers during a meeting with Wall Street analysts. Following this, the stocks of early quantum computer development companies, including IonQ, plummeted intraday, causing turmoil in the sector.

The quantum computing industry has shown various reactions to Huang's remarks, ranging from direct rebuttals to indirect criticisms. However, before hastily defining the meaning of Huang's statement, it is necessary to look back on the life he has walked so far. His ambition with 'GPGPU,' which has now borne fruit in AI semiconductors, also took exactly 20 years.

NVIDIA’s Cash Cow Dominating the AI Market, GPGPU

The fundamental difference between GPGPU and gaming GPUs lies in the software. GPGPU enhances the bandwidth between the GPU, memory chips, and CPU to deliver processing power specialized for high-performance computing and AI computing. Screenshot from NVIDIA's official website

NVIDIA’s graphics processing units (GPUs) currently dominate 90% of the AI semiconductor market. Originally, GPUs were used as semiconductors for 3D game graphics processing, but now they are installed in data centers and supercomputers. However, AI GPUs and gaming GPUs are actually different. The origin of the data center GPUs NVIDIA sells today traces back to the 'GPGPU' development project that began in the early 2000s.

GPGPU stands for General-Purpose GPU. In other words, it refers to computer chips that extend the GPU’s tasks beyond general graphics processing to other computing functions. Considering that Huang’s NVIDIA grew closely tied to the gaming industry, especially the game console industry, this was a very bold attempt.

Initially Met with Cold Reception... Jensen Huang Also Took 20 Years

The U.S. Defense Advanced Research Projects Agency (DARPA) collaborated with American semiconductor companies such as Nvidia and Intel in 2010 to conduct the 'Ubiquitous High Performance Computing' research project. Although the tangible commercialization of this project’s outcomes has never occurred, it indirectly nourished various semiconductor innovations, including GPGPU. DARPA

What market was Huang targeting with GPGPU development? In the 2000s, when the GPGPU architecture was being developed, machine learning (ML) was still an unfamiliar term not only to the general public but also to investors. Huang did not foresee the AI boom at that time.

Instead, Huang anticipated a major transformation in the science, medical, and financial industries through GPGPU. For example, GPGPU could be used instead of traditional CPUs for simulation tasks related to physics, thermodynamics, and biology. It was also expected to be applied in processing large volumes of medical images and diagnostic device data or in computational finance such as algorithmic trading. NVIDIA even challenged next-generation military parallel computing chip research in collaboration with the U.S. Defense Advanced Research Projects Agency (DARPA) in 2010.

NVIDIA has not always been on a winning streak. Its revenue stagnated for several years following the 2008 financial crisis.

However, as the past decade of history has shown, the true potential of GPGPU was neither in simulations, finance, nor defense industries. In fact, NVIDIA’s annual revenue was practically stagnant from 2007 to 2015, and during this period, Huang’s GPGPU project faced much criticism. Until the AI boom triggered by Google DeepMind’s deep learning and OpenAI’s large language models (LLMs) 'rescued' NVIDIA, Huang searched for GPGPU demand across numerous industries such as gaming, general graphics, and cryptocurrency mining.

Technology Alone Is Not Enough... Finding the Right Market Is Key

Jensen Huang, CEO, presenting on the future of GPUs at the 2013 GTC conference. Photo by NVIDIA YouTube capture

In this sense, Huang’s '20 years' remark is more likely to refer to 'demand' rather than the 'feasibility' of quantum computers. In fact, NVIDIA is probably one of the companies that knows the trends in quantum computer technology development better than anyone else.

NVIDIA’s GPGPU already provides a quantum computing acceleration software development kit (SDK) called 'cuQuantum.' NVIDIA has worked with numerous quantum computer startups and big tech companies and continues to maintain partnerships.

However, there is a big difference between quantum computers being mass-produced at a feasible scale and being actually useful in the market. Currently, the computational capabilities leveraging the unique properties of quantum computers are expected to be useful in areas such as computational cryptography, biological simulations, and data traffic management, but these markets are still small and can be sufficiently handled by conventional computers.

Of course, in the near future, the price-to-performance ratio of quantum computers may improve at an astonishing rate, potentially entering a phase similar to what GPGPU experienced starting in 2016. But any innovative technology must go through many trials and tribulations before settling into the 'right' market, and as Huang said, it might indeed take 20 years.