Personal blog of Yzmir Ramirez

Artificial intelligence technology holds a huge amount of promise for enterprises — as a tool to process and understand their data more efficiently; as a way to leapfrog into new kinds of services and products; and as a critical stepping stone into whatever the future might hold for their businesses. But the problem for many enterprises is that they are not tech businesses at their core, so bringing on and using AI will typically involve a lot of heavy lifting. Today, one of the startups building AI services is announcing a big round of funding to help bridge that gap.

SambaNova — a startup building AI hardware and integrated systems that run on it that only officially came out of three years in stealth last December — is announcing a huge round of funding today to take its business out into the world. The company has closed on $676 million in financing, a Series D that co-founder and CEO Rodrigo Liang has confirmed values the company at $5.1 billion.

The round is being led by SoftBank, which is making the investment via Vision Fund 2. Temasek and the government of Singapore Investment Corp. (GIC), both new investors, are also participating, along with previous backers BlackRock, Intel Capital, GV (formerly Google Ventures), Walden International and WRVI, among other unnamed investors. (Sidenote: BlackRock and Temasek separately kicked off an investment partnership yesterday, although it’s not clear if this falls into that remit.)

Co-founded by two Stanford professors, Kunle Olukotun and Chris Ré, and Liang, who had been an engineering executive at Oracle, SambaNova has been around since 2017 and has raised more than $1 billion to date — both to build out its AI-focused hardware, which it calls DataScale, and to build out the system that runs on it. (The “Samba” in the name is a reference to Liang’s Brazilian heritage, he said, but also the Latino music and dance that speaks of constant movement and shifting, not unlike the journey AI data regularly needs to take that makes it too complicated and too intensive to run on more traditional systems.)

SambaNova on one level competes for enterprise business against companies like Nvidia, Cerebras Systems and Graphcore — another startup in the space which earlier this year also raised a significant round. However, SambaNova has also taken a slightly different approach to the AI challenge.

In December, the startup launched Dataflow-as-a-Service as an on-demand, subscription-based way for enterprises to tap into SambaNova’s AI system, with the focus just on the applications that run on it, without needing to focus on maintaining those systems themselves. It’s the latter that SambaNova will be focusing on selling and delivering with this latest tranche of funding, Liang said.

SambaNova’s opportunity, Liang believes, lies in selling software-based AI systems to enterprises that are keen to adopt more AI into their business, but might lack the talent and other resources to do so if it requires running and maintaining large systems.

“The market right now has a lot of interest in AI. They are finding they have to transition to this way of competing, and it’s no longer acceptable not to be considering it,” said Liang in an interview.

The problem, he said, is that most AI companies “want to talk chips,” yet many would-be customers will lack the teams and appetite to essentially become technology companies to run those services. “Rather than you coming in and thinking about how to hire scientists and hire and then deploy an AI service, you can now subscribe, and bring in that technology overnight. We’re very proud that our technology is pushing the envelope on cases in the industry.”

To be clear, a company will still need data scientists, just not the same number, and specifically not the same number dedicating their time to maintaining systems, updating code and other more incremental work that comes managing an end-to-end process.

SambaNova has not disclosed many customers so far in the work that it has done — the two reference names it provided to me are both research labs, the Argonne National Laboratory and the Lawrence Livermore National Laboratory — but Liang noted some typical use cases.

One was in imaging, such as in the healthcare industry, where the company’s technology is being used to help train systems based on high-resolution imagery, along with other healthcare-related work. The coincidentally-named Corona supercomputer at the Livermore Lab (it was named after the 2014 lunar eclipse, not the dark cloud of a pandemic that we’re currently living through) is using SambaNova’s technology to help run calculations related to some COVID-19 therapeutic and antiviral compound research, Marshall Choy, the company’s VP of product, told me.

Another set of applications involves building systems around custom language models, for example in specific industries like finance, to process data quicker. And a third is in recommendation algorithms, something that appears in most digital services and frankly could always do to work a little better than it does today. I’m guessing that in the coming months it will release more information about where and who is using its technology.

Liang also would not comment on whether Google and Intel were specifically tapping SambaNova as a partner in their own AI services, but he didn’t rule out the prospect of partnering to go to market. Indeed, both have strong enterprise businesses that span well beyond technology companies, and so working with a third party that is helping to make even their own AI cores more accessible could be an interesting prospect, and SambaNova’s DataScale (and the Dataflow-as-a-Service system) both work using input from frameworks like PyTorch and TensorFlow, so there is a level of integration already there.

“We’re quite comfortable in collaborating with others in this space,” Liang said. “We think the market will be large and will start segmenting. The opportunity for us is in being able to take hold of some of the hardest problems in a much simpler way on their behalf. That is a very valuable proposition.”

The promise of creating a more accessible AI for businesses is one that has eluded quite a few companies to date, so the prospect of finally cracking that nut is one that appeals to investors.

“SambaNova has created a leading systems architecture that is flexible, efficient and scalable. This provides a holistic software and hardware solution for customers and alleviates the additional complexity driven by single technology component solutions,” said Deep Nishar, senior managing partner at SoftBank Investment Advisers, in a statement. “We are excited to partner with Rodrigo and the SambaNova team to support their mission of bringing advanced AI solutions to organizations globally.”

Superlatives abound at Cerebras, the until-today stealthy next-generation silicon chip company looking to make training a deep learning model as quick as buying toothpaste from Amazon. Launching after almost three years of quiet development, Cerebras introduced its new chip today — and it is a doozy. The “Wafer Scale Engine” is 1.2 trillion transistors (the most ever), 46,225 square millimeters (the largest ever), and includes 18 gigabytes of on-chip memory (the most of any chip on the market today) and 400,000 processing cores (guess the superlative).

It’s made a big splash here at Stanford University at the Hot Chips conference, one of the silicon industry’s big confabs for product introductions and roadmaps, with various levels of oohs and aahs among attendees. You can read more about the chip from Tiernan Ray at Fortune and read the white paper from Cerebras itself.

Superlatives aside though, the technical challenges that Cerebras had to overcome to reach this milestone I think is the more interesting story here. I sat down with founder and CEO Andrew Feldman this afternoon to discuss what his 173 engineers have been building quietly just down the street here these past few years, with $112 million in venture capital funding from Benchmark and others.

Going big means nothing but challenges

First, a quick background on how the chips that power your phones and computers get made. Fabs like TSMC take standard-sized silicon wafers and divide them into individual chips by using light to etch the transistors into the chip. Wafers are circles and chips are squares, and so there is some basic geometry involved in subdividing that circle into a clear array of individual chips.

One big challenge in this lithography process is that errors can creep into the manufacturing process, requiring extensive testing to verify quality and forcing fabs to throw away poorly performing chips. The smaller and more compact the chip, the less likely any individual chip will be inoperative, and the higher the yield for the fab. Higher yield equals higher profits.

Cerebras throws out the idea of etching a bunch of individual chips onto a single wafer in lieu of just using the whole wafer itself as one gigantic chip. That allows all of those individual cores to connect with one another directly — vastly speeding up the critical feedback loops used in deep learning algorithms — but comes at the cost of huge manufacturing and design challenges to create and manage these chips.

The first challenge the team ran into, according to Feldman, was handling communication across the “scribe lines.” While Cerebras’ chip encompasses a full wafer, today’s lithography equipment still has to act like there are individual chips being etched into the silicon wafer. So the company had to invent new techniques to allow each of those individual chips to communicate with each other across the whole wafer. Working with TSMC, they not only invented new channels for communication, but also had to write new software to handle chips with trillion-plus transistors.

The second challenge was yield. With a chip covering an entire silicon wafer, a single imperfection in the etching of that wafer could render the entire chip inoperative. This has been the block for decades on whole-wafer technology: due to the laws of physics, it is essentially impossible to etch a trillion transistors with perfect accuracy repeatedly.

Cerebras approached the problem using redundancy by adding extra cores throughout the chip that would be used as backup in the event that an error appeared in that core’s neighborhood on the wafer. “You have to hold only 1%, 1.5% of these guys aside,” Feldman explained to me. Leaving extra cores allows the chip to essentially self-heal, routing around the lithography error and making a whole-wafer silicon chip viable.

Entering uncharted territory in chip design

Those first two challenges — communicating across the scribe lines between chips and handling yield — have flummoxed chip designers studying whole-wafer chips for decades. But they were known problems, and Feldman said that they were actually easier to solve than expected by re-approaching them using modern tools.

He likens the challenge to climbing Mount Everest. “It’s like the first set of guys failed to climb Mount Everest, they said, ‘Shit, that first part is really hard.’ And then the next set came along and said ‘That shit was nothing. That last hundred yards, that’s a problem.’ ”

And indeed, the toughest challenges, according to Feldman, for Cerebras were the next three, since no other chip designer had gotten past the scribe line communication and yield challenges to actually find what happened next.

The third challenge Cerebras confronted was handling thermal expansion. Chips get extremely hot in operation, but different materials expand at different rates. That means the connectors tethering a chip to its motherboard also need to thermally expand at precisely the same rate, lest cracks develop between the two.

As Feldman explained, “How do you get a connector that can withstand [that]? Nobody had ever done that before, [and so] we had to invent a material. So we have PhDs in material science, [and] we had to invent a material that could absorb some of that difference.”

Once a chip is manufactured, it needs to be tested and packaged for shipment to original equipment manufacturers (OEMs) who add the chips into the products used by end customers (whether data centers or consumer laptops). There is a challenge though: Absolutely nothing on the market is designed to handle a whole-wafer chip.

“How on earth do you package it? Well, the answer is you invent a lot of shit. That is the truth. Nobody had a printed circuit board this size. Nobody had connectors. Nobody had a cold plate. Nobody had tools. Nobody had tools to align them. Nobody had tools to handle them. Nobody had any software to test,” Feldman explained. “And so we have designed this whole manufacturing flow, because nobody has ever done it.” Cerebras’ technology is much more than just the chip it sells — it also includes all of the associated machinery required to actually manufacture and package those chips.

Finally, all that processing power in one chip requires immense power and cooling. Cerebras’ chip uses 15 kilowatts of power to operate — a prodigious amount of power for an individual chip, although relatively comparable to a modern-sized AI cluster. All that power also needs to be cooled, and Cerebras had to design a new way to deliver both for such a large chip.

It essentially approached the problem by turning the chip on its side, in what Feldman called “using the Z-dimension.” The idea was that rather than trying to move power and cooling horizontally across the chip as is traditional, power and cooling are delivered vertically at all points across the chip, ensuring even and consistent access to both.

And so, those were the next three challenges — thermal expansion, packaging and power/cooling — that the company has worked around-the-clock to deliver these past few years.

From theory to reality

Cerebras has a demo chip (I saw one, and yes, it is roughly the size of my head), and it has started to deliver prototypes to customers, according to reports. The big challenge, though, as with all new chips, is scaling production to meet customer demand.

For Cerebras, the situation is a bit unusual. Because it places so much computing power on one wafer, customers don’t necessarily need to buy dozens or hundreds of chips and stitch them together to create a compute cluster. Instead, they may only need a handful of Cerebras chips for their deep-learning needs. The company’s next major phase is to reach scale and ensure a steady delivery of its chips, which it packages as a whole system “appliance” that also includes its proprietary cooling technology.

Expect to hear more details of Cerebras technology in the coming months, particularly as the fight over the future of deep learning processing workflows continues to heat up.

 While some of the largest chip manufacturers are looking to shift their focus onto the GPU for their biggest machine learnings, there’s a blooming ecosystem of new chip startups looking to rethink the way processing for AI works That includes a European-based startup called Graphcore, which said today that it has raised $50 million in new financing led by Sequoia Capital — following… Read More

 Benchmark is doing another big investment before the end of the year, and this time it’s in a hardware startup called Cerebras Systems. From what we hear from sources, the round (which we heard others were participating in) is somewhere around $25 million, though it could be as high as $30 million. Details on the company are extremely sparse — we can’t even really find… Read More