Quantum Hype Keeps Missing the Real Big Bang
The quantum computing hype just keeps on coming. Only briefly deflating the hypesters were the wise words of Jensen Huang, CEO of Nvidia (which is running a well-funded quantum computing research project of its own), who explained it would be 20 or 30 years before quantum computing becomes significant. That is the way technologists say, “never.”
Then Microsoft announced its “Majorana 1” chip and the hype burst its banks again. Forbes gushed that “the race toward scalable quantum computing has reached a pivotal moment… Majorana 1… marks a significant milestone…. As the quantum computing industry debates the timeline for practical implementation, breakthroughs like Majorana 1 and Google’s Willow suggest that major advancements may be closer than previously thought.”
Even better, quantum computing is “diverse” because Willow, Majorana, and other projects such as “IBM’s long-term quantum roadmap illustrate the industry’s diverse approaches to achieving fault-tolerant quantum systems.” Read “diverse” as “lacking a coherent theory and making many stabs in the darkness of quantum nirvana.” Read “fault tolerant” as “it doesn’t work but we’ve got the funding.”
Dr. Ekta Dang, author of the Forbes lyric, says Majorana is a breakthrough because it “aims to build fault tolerance directly into the hardware.” Satya Nadella, Microsoft’s CEO, proclaims, “We’ve created an entirely new state of matter, powered by a new class of materials, topo-conductors,” which bring the path to “a million-qubit processor within reach—bringing us closer to solving problems beyond the capabilities of classical computing.”
Meanwhile, Google’s Willow chip is “designed to exponentially reduce errors as more qubits are added… by leveraging advanced error correction.” What really makes the headlines, though, is that in benchmarking tests “Willow performed a computation in under five minutes that would take one of today’s fastest supercomputers an estimated 10 septillion years.”
Yes, but how long does it take to ask a useful question that takes 10 septillion years to answer?
That’s the relevant issue because the quantum computer is really just a hugely complex and elusive analog machine. It mimics reality rather than capturing it or interpreting it. It is based on inchoate quantum theories that have not even been proven to be true explanations of quantum phenomena. For example, quantum theory completely failed to anticipate the coherence of 2D graphene; hence the five Nobel Prizes for topological phases of matter that Microsoft is now exploiting.
To translate problems into the elusive analog superpositions of qubits seems mostly a step backward. The Microsoft and Google machines are, at best, analog physics simulators that still don’t really work. When you’re through, you still have the input and output to master and measure. And to make any use of it, you will need to translate the answer to your 10-septillion-year question to the conventional digital language that made computers possible in the first place. Do you really have another 10 septillion years on your hands?
Companies that pretend to have great futures because of their quantum computing pretensions are chiefly just preening in the limelight of gullible media.
On the other hand, all computers are based in part on discrete applications of quantum theory, with their tunnel diodes, and Josephson junctions, and other functions. But computers work because they are digital, and thus, they abstract from the quantum mystagogy.
Meanwhile, two real and far more consequential revolutions in computation are already at hand.
No more chips, we’re makin’ wafers!
The transition from chips to wafer-scale computing is a current reality, led by Cerebras. Its third-generation computing engine is 12 inches square, or 46,225 mm2, with four trillion transistors and 900,000 AI “cores”. (For a rough comparison, if you have a top-of-the-line MacBook Pro, the chip that drives it is about 400mm2, with 28 billion transistors.)
The Cerebras machine can execute at 125 petaflops. That’s 125 quadrillion floating-point operations per second. It can train AI models up to 24 trillion parameters.
AI was made practical by connecting thousands of Nvidia chips to solve problems in parallel. A magnificent achievement. But connections across chips working in parallel are massive energy wasters and heat generators; the chief reason hyperscale data centers now use as much energy for cooling as they do for computing.
The energy saved, heat avoided, and speed gained by staying on a single chip is the reason the industry has so passionately pursued the Moore’s law standard of periodically doubling the number of transistors on a single chip. Practically speaking, that path to progress is now closed.
The potential payoff for going to wafer scale has always been huge but plagued by the speed-of-light limits on reticle size for extreme ultraviolet lithography from ASML in Holland. Then, there are those “error correction” issues. That’s because the probability of a defect on a chip or a wafer is a function of the area and the number of circuits. It took Cerebras almost a decade to solve the problem. Cerebras devices will eventually put a data center on your lap.
Graphene chills…
Impressive? You ain’t seen nothin’ yet. The next big bang will come in two stages. First is the migration from copper and silver for interconnects to graphene, with 1000 times the conductivity of copper. Graphene connectors could render the current data center heat crisis trivial, creating hundreds of billions in value. We will have workable graphene connectors not in 20 years (i.e., never) but within this decade. In this week’s issue of our sister publication, Gilder’s Private Reserve, we profile a company poised to deliver. And we show our readers how to get in on the deal with a minimum investment of just $250!!!
… and thrills…
Beyond replacing copper on and between chips, the next breakthrough will come as graphene replaces silicon itself, accelerating computation at a pace equivalent to decades of Moore’s law overnight. Unlike the quantum computer, the physics of wafer-scale graphene computation has already been demonstrated by a team drawn from Georgia Tech and Tianjin University (China’s oldest) and led by a Dutchman, Walter de Heer, Regents’ Professor of Physics at Georgia Tech.
… and leaves quantum standing still
The wafer-scale, graphene supercomputer will operate not at petaflops or exaflops (quintillions), but zettaflops, sextillions (1023) of operations per second, and use less energy than silicon machines running a million times more slowly. And all those operations will be digital, rendering the quantum project a costly curiosity.