Sizeable Series C round brings Groq's total funding to $367 million, yet it is nonetheless dwarfed by competitors' raises in this space.
Following the recent funding announcement from competitor SambaNova, data center AI chip startup Groq has announced completion of a Series C funding round of $300 million. This brings Groq’s total funding to $367 million, a sizeable total, but still less total funding to date than key startup competitors SambaNova ($1.1 billion) and Graphcore ($710 million), though those companies have completed D and E rounds, respectively.
“We were aiming for a much more modest sum than this,” Jonathan Ross, CEO of Groq told EE Times. “However, as investors began speaking with our customers, they got very excited and decided to invest more. We actually had to push back on additional funding beyond this.”
Series C funding of $300 million has all been raised in the second half of 2020, following customers’ reviews of first generation Groq hardware, according to Ross.
“This is more than we imagined raising, but we know exactly how to apply it,” Ross said, noting that Groq is planning to expand its team of 122 to 250 by the end of the year. The company currently has 43 positions advertised on its website.
“Historically, there’s been a flow of talent out of hardware and into software companies,” Ross said. “If you’ve got an EE degree, what it usually meant was you were going to be at a hardware company for a few years before you went to one of the exciting software companies. And what we’re now seeing is a reversal of that trend… the companies that we are hiring people from, a lot of them are very well-known software companies. Some of these folks have EE degrees, but not all of them — many of them are pure software engineers. Hardware is the new software – it’s become exciting again.”
As well as focusing on talent, a second-generation tensor streaming processor (TSP) architecture and product is “pretty far along,” said Ross. “The money will help us scale it and take it to production. But also other initiatives that we’re working on as well, that will come a little earlier, we’ll now be able to take those to production as well.”
Groq positions its TSP architecture as the epitome of simplicity – a single giant core up against GPUs’ thousands of cores, simple to program, with fast and predictable latency (to the nanosecond). This predictable latency scales to multi-chip implementations since Groq’s TSP has sufficient I/O bandwidth to allow hundreds of Groq chips to work synchronously. A program that takes one chip 43,000 nanoseconds to complete will take 430 nanoseconds on a hundred chips, Ross said.
“It will take that amount of time, plus or minus zero, every time, even though it’s operating across a hundred chips,” Ross said.
Groq’s low latency is appealing to the autonomous vehicle (AV) sector, while throughput is more important to data centers. Ross argues that Groq’s TSP is a fit for both, though the applications are totally different. AVs might run 30-40 distinct computer vision models on the same hardware, whereas a financial institution might run a single problem, 30-40 lines of code, at huge scale. Groq is approaching these two very different markets with the same chip.
“We actually have lighthouse customers in AVs and data centers using the exact same chip, but they use the chip for very different purposes,” said Ross. “As an example in the AV situation, they’re taking one of our chips and displacing four high-end GPUs. And in the data center case, they’re actually connecting hundreds of our chips together in order to be able to solve problems more quickly.”
Groq’s Series C funding round was co-led by Tiger Global Management and D1 Capital, with participation from The Spruce House Partnership and Addition.