Belgium’s Vertical Compute Raises €20M Seed to Tackle AI Chips’ Memory Bottleneck
-
Loading... - Olive Rivera
- 31 Jan 2025
- 35 Views
- 0 Like
- 0 Comment
Vertical Compute, a Belgian startup focused on developing more efficient chips for artificial intelligence (AI), has successfully raised €20 million in an all-equity seed round just three months after its launch. The company was co-founded by Sylvain Dubois, a former Google semiconductor strategist, and Sebastien Couet, who led semiconductor programs at imec, a renowned R&D center in Leuven, Belgium.
The funding round was led by imec.xpand, an independent venture capital fund that invests in startups leveraging technologies developed at imec. Other investors included French venture capital firms Eurazeo and XAnge, as well as Swiss investor Vector Gestion.
Vertical Compute aims to address the growing "memory bottleneck" in AI, a significant challenge as AI models rapidly increase in scale, outpacing the capabilities of existing memory technologies. The market for AI chips is substantial, valued at approximately $150 billion, with major players like Samsung, Micron, and Hynix dominating the sector.
Addressing the Memory Bottleneck
Current AI chips, such as Nvidia's graphics processing units (GPUs), rely on external high-bandwidth memory (HBM) components, primarily produced by the aforementioned companies. Dubois explains that this architecture requires data to be stored separately from the processing unit, leading to inefficiencies. Each chip can incur around $10,000 in memory costs due to the expensive HBMs.
Vertical Compute is developing a new chip architecture that integrates memory layers directly into the chip itself. This innovation aims to reduce data movement from centimeters to nanometers, potentially outperforming traditional HBMs in terms of density, cost, and energy efficiency. By having data readily available on the chip, the startup anticipates significant gains in energy consumption and infrastructure costs.
Enabling Edge Computing
One of the primary applications for Vertical Compute's technology is in edge computing, where computations occur on local devices—such as laptops and smartphones—rather than relying on centralized cloud data centers. Currently, AI calculations are predominantly cloud-based due to the high costs and complexities associated with integrating HBMs into consumer devices.
Dubois argues that the reliance on cloud architectures, which are primarily developed by a few tech giants like Google, Amazon, and Microsoft, is inefficient in terms of energy, latency, and data protection. Vertical Compute's chips are designed for inference tasks, allowing pre-trained AI models to apply their knowledge to new data locally, thus reducing the need for constant cloud interaction.
Future Plans and Market Strategy
Vertical Compute does not intend to manufacture chips entirely on its own. Instead, the startup will design the additional memory layers and create test products in-house, outsourcing large-scale production to semiconductor foundries. The technology is expected to be compatible with products made by Taiwan Semiconductor Manufacturing Company (TSMC), a global leader in chip manufacturing.
The startup plans to target semiconductor designers like Nvidia, AMD, and Intel, who currently depend on traditional HBM suppliers. Although no partnerships have been finalized, Dubois notes that Nvidia has identified Micron as a key memory component supplier.
Despite its international ambitions, Vertical Compute aims to remain a European company, emphasizing the need for such enterprises in Europe. Dubois, who returned to Europe after 25 years in the US, plans to recruit around 30 employees over the next year, focusing on talent in Belgium and France, where there are strong semiconductor hubs.
In conclusion, Vertical Compute's innovative approach to tackling the memory bottleneck in AI chips positions it as a promising player in the semiconductor industry, with the potential to significantly impact the future of AI technology.
