In March, Nvidia announced its advanced Blackwell GPU platform, in a strategic move aimed at enhancing AI processing capabilities and reinforcing its market position. Watching from the sidelines and encouraged by the announcement was Elad Raz, CEO of Israeli company NextSilicon a rival of Nvidia in the chip market.
"The Blackwell announcement was the best thing that could happen to us," Raz says. "Customers who own supercomputers don’t need to run AI applications and therefore don’t need such expensive chips. They felt that having to buy Blackwell would be disastrous for them. It consumes twice as much power as the current chip, costs twice as much, and provides a similar level of performance. It was a moment of disillusionment for our customers, who realized that building a new supercomputer would now cost them four times more than before."
Don't call it a processor
Over its seven years of operation, NextSilicon has worked on launching a chip it claims will beat the AI market. It has so far raised $300 million, at a $1.5 billion valuation, from venture capital funds such as billionaire Dan Loeb's Third Point Ventures, Standard Investments, Playground Global, Amiti Ventures, and Aleph VC. So far, it has launched an experimental chip, "Maverick-1", which it has sold to universities, research institutes, and US government agencies for tens of millions of dollars, to test the feasibility of the concept.
It is now launching a high-performance chip for sale on the open market called Maverick 2, which claims to have the highest performance-to-power-consumption ratio. However, Raz is careful not to call the chip - developed in Israel and manufactured by TSMC in Taiwan - a "processor", preferring instead to call it an "Intelligent Compute Accelerator." "It's a completely different thing. The chip accelerates applications four times faster than graphics processors and 20 times faster than core processors, and the basis for this is a combination of hardware and software," he explains.
The war broke out just before the "tape-out" release date - the stage when chips are sent for production - forcing the company to delay for two months. "On October 8, we started inscribing 'Am Yisrael Chai' on every Maverick-2 chip, which will go to each of our customers, and live there forever," says Raz.
Before founding NextSilicon in 2017, Raz served as director of software at Mellanox, after Mellanox’s acquisition of his software company, Integrity Project, for $10 million. During his tenure, he wondered why chips had been developed with the same hardware structure for decades, when integrating software directly with the hardware could potentially improve performance. According to Raz, "Everything that exists on the market today - from graphics processors, through core processors or accelerators, to AI or machine learning - are processors that receive commands. We use an algorithm based on the Pareto principle, or the 20/80 rule, to identify what is most important at any given moment for running the software, performing calculations, and reconfiguring the wiring in the hardware."
In essence, Raz explains, the algorithm instantly designs a temporary, dedicated chip in the hardware to run the specific software. "In this way, electricity utilization is more efficient because only a small part of the chip is used. In a normal processor, a small percentage [of components] perform mathematical operations, while the rest are engaged in data transfer. Ours is more half and half."
How unique is it compared to what exists today? Do you have any rivals?
"We are the only ones with a chip that isn't a processor. Any company aiming to compete with us will need at least four years of work, starting from scratch, to write all the hardware and software."
How to beat Nvidia
Do you expect major companies to replace processors from industry giant Nvidia with those from an Israeli startup based in Givatayim?
"I don’t expect us to have companies lining up tomorrow or to earn $26 billion a quarter like Nvidia. Developing Blackwell cost them $10 billion - how can we compete when we’ve invested just a fraction of that, around $100 million? Our advantage lies in the unique architecture we've developed. You can’t beat Nvidia with more of the same. We believe our chip is more efficient and has the potential to be a winner in the AI and database sectors. For now, we’re starting where we have a clear advantage: supercomputers, where we help large organizations answer questions that the current technology can't address. Once we've proven ourselves in that market, we’ll expand into the AI space."
NextSilicon is currently working with government research institutes and universities in building local computers capable of handling complex operations, such as simulating molecular interactions in drug development, modeling engine performance, and generating energy through fusion processes in laboratory settings. The company works with the US Department of Defense as well as with Sandia National Laboratories, a contractor for the US Department of Energy’s National Nuclear Security Administration, that next year will supply a supercomputer, called "Spectra," which will be based on the Maverick-2.
You’ve chosen a niche market, not one that will bring you global distribution.
"This is the market that will change the world of computing and electric circuits. The market generates $50 billion every year, of which $20 billion are in chip sales. If that’s niche, it's not at all small."
The company is already preparing to enter the AI server market and other applications in 2026, both in training models and their operation. "By 2026, we will launch a chip that is three times better than Nvidia's. Moreover, while no one knows what language models will look like in the future, our chip is designed to adapt to any technology, enabling it to support models that haven't yet been created."
Nvidia has complete control over the entire AI market value chain, from processors to manufacturers like Dell and Super Micro, who build their servers based on Nvidia's dictates. This is a level of power and influence you don't have.
"I agree that we don't have the same power as Nvidia to tailor supercomputers for large organizations. It has processors, servers, and communication components purchased from Mellanox for billions of dollars, and it can provide customers with everything they need. We're not yet able to invest billions in building a vertical arm like that, which is why we’re initially turning to more high-level customers. This is our disadvantage compared to Nvidia."
Military AI training
Some say Israel has lost its lead in AI, or that it never had it, unlike its lead in the cybersecurity and software sectors. Raz says, "The Israeli ecosystem is geared towards cybersecurity, which is rooted in security needs. High schools offer a course in cybersecurity but no course in neural networks. In my opinion, it is a mandatory subject, certainly in the army."
"The army must begin investing in training more soldiers in AI, neural networks, and computing infrastructure. Future wars will demand these skills. What will we do with all the intelligence data we gather? We need systems that can make connections, draw associations, and anticipate enemy movements. The lesson of October 7 mandates this."
The state budget for all AI programs is $500 million for three years, a tiny amount compared to investment in the US, China, or the UK.
"If we look around us, all advanced countries have national supercomputers. Israel doesn't have one; we use other countries’ computers. The state must understand that computing infrastructure is a strategic asset. The moment a supercomputer is constructed here on a national scale, technologies that are yet to be discovered will emerge from here."
Israel no longer establishes chip companies the way it did in the past.
"This is an international problem. The cost of developing a chip starts in the region of $120-150 million. To raise an amount like that when founding a company, you need a really special story that describes how you will generate revenue, and explains how you will compete with giants like Nvidia. It's not easy, but companies like that did come up in 2020-2021. In 2023, the flow of investment stopped."
Published by Globes, Israel business news - en.globes.co.il - on November 3, 2024.
© Copyright of Globes Publisher Itonut (1983) Ltd., 2024.