Vertical Compute, a Belgian startup that is developing more efficient chips for AI, has raised an all-equity €20m seed round just three months after launching.
The startup was cofounded by Sylvain Dubois, who previously spent four years working on Google’s semiconductor strategy in San Francisco, and Sebastien Couet, who for the past 10 years led semiconductor programs at imec, an R&D centre based in Leuven in Belgium, which focuses on nanotechnologies.
The startup is an imec spinout and the round was led by the research centre’s VC branch imec.xpand. It also included French VCs Eurazeo and XAnge, as well as Swiss investor Vector Gestion.
Vertical Compute is developing a new architecture for chips that power AI models, which it says will overcome the ‘memory bottleneck’ — a fast-growing issue in the sector, due to the fact that AI models are rapidly increasing in scale and outpacing the capabilities of the memory technologies that they need to function.
It’s a huge market — worth $150bn, Dubois tells Sifted. “And three companies own it: Samsung, [US semiconductor giant] Micron and [Korean semiconductor company] Hynix.”
“So, we’re going to tickle trillion-dollar companies.”
The memory bottleneck
The chips that power AI, such as Nvidia’s graphics processing units (GPUs), are currently reliant on external memory components called high-bandwidth memories (HBM), which are mostly produced by Samsung, Micron and Hynix.
“You can’t store data in the same component as where the calculation happens,” says Dubois.
Instead, data is stored in HBMs and fed into the processor to enable computations. “HBMs sit on top of the chip and are extremely costly — for every chip, you’ve got about $10k of memory,” says Dubois.
Vertical Compute is developing an architecture that it says will integrate layers of memory directly into the chip. The company says that reducing data movements from centimeters to nanometers has the potential to outperform regular HBMs in terms of density, cost and energy.
“The data is available directly,” says Dubois. “The component is there, there are no back-and-forths. So you gain energy, infrastructure and thousands of dollars spent on memory chips.”
Enabling edge computing
One of the key use cases that will be boosted thanks to the new technology, says Dubois, is edge computing — when computations happen on local devices, such as laptops and phones, rather than through the cloud in centralised data centres.
Currently, AI calculations are largely cloud-based. Dubois says that this is because HBMs are too costly and complex to integrate into users’ devices, and therefore sit in data centres.
The back-and-forths between cloud infrastructure and devices, however, mean data transfer latency, higher energy consumption and more costs — especially as AI models grow in scale and require more data to be processed quickly.
“It makes no sense to depend on cloud architectures that only three companies can afford to develop — Google, Amazon and Microsoft — and have every AI request we make sent to the cloud, processed, and sent back,” says Dubois. “It makes no sense energy-wise, latency-wise, or in terms of data protection. We need an alternative.”
While Vertical Compute’s chips isn’t intended to train AI models entirely on a local device, they will be used for inference, says Dubois — the stage where a pre-trained model applies its knowledge to new data for a specific use case.
Selling to Nvidia
Vertical Compute doesn’t plan to manufacture new chips entirely. The startup will design the additional layers of memory and make test products in-house, and then outsource industrial production to semiconductor foundries.
Although no partnership has been signed yet, Dubois says that the company’s technology will be compatible with the products made by Taiwan-based TSMC, a global leader in chip manufacturing.
Vertical Compute will be looking to sell its technology to semiconductor designers like Nvidia, AMD or Intel, which currently rely on traditional HBM manufacturers. Nvidia recently said that Micron is its key memory component supplier.
Although the startup’s key supplier is likely to be in Taiwan and biggest customers in the US, Dubois says that Vertical Compute is planning to remain a European company.
“There is a huge need for Europe to have this type of company,” he says. “We have international ambitions, but we want to create a European company, with European talent and resources, to give back to Europe after 25 years working for American companies.”
Dubois came back from the US six months ago to launch Vertical Compute. He is planning to recruit around 30 people in the next year, in Belgium and in France — where he says there are strong semiconductor hubs in Grenoble and Nice. A French subsidiary is scheduled to open in the next few weeks.
And to grow his team, Dubois is looking across the Atlantic. “I’ve experienced the exodus [of Europeans] to California and seen the mass return since Covid. Repatriating talent is something I am very keen on — and I think there are still a few people to convince in California.”
Read the orginal article: https://sifted.eu/articles/vertical-compute-20m-seed-news/
