Back
14.01.2025

Transforming the future of computing: why we invested in Vertical Compute?

By Julie Forel

In the midst of an artificial intelligence gold rush, we are excited to announce our co-lead investment in Vertical Compute’s €20 millions seed round. Founded in 2024, Vertical Compute aims to revolutionize the memory industry and become a world leader in this critical sector.

I – The AI Revolution and the Memory Wall

The current AI boom has catapulted equipment providers to unprecedented heights, with Nvidia leading the charge. As Nvidia’s market capitalization surpasses $3 trillion, it’s clear that the value of AI infrastructure cannot be overstated.

At the heart of this technological boom lies the Graphics Processing Unit (GPU), a specialized chip crucial for AI applications. GPUs are designed to swiftly perform complex parallel calculations, primarily for images and videos, and now for AI. These powerhouses consist of two main components:

  • the processor, which handles processing tasks,
  • the memory, which efficiently feeds the processor with data.

As AI continues to advance, the demand for more powerful and efficient memory solutions grows exponentially. This is where Vertical Compute enters the picture, with an ambitious strategy to create a unique memory solution that could potentially overcome the limitations of current technologies like DRAM and SRAM.

 

II – The Current Memory Market Landscape

Main players in the Memory Market

The memory market is currently dominated by established players like Samsung, SK Hynix, and Micron. These companies lead in DRAM and NAND flash production but face challenges in scaling and cost reduction. The industry has consolidated significantly, with the top 3 suppliers now owning more than 95% of the market, down from over 20 manufacturers in the mid-1990s.

 

The Main DRAM variants

DRAM comes in different variants, each optimized for different goals. The differences between them lie almost entirely in the peripheral circuits. The memory cells themselves are similar across varieties and the fabrication methods are broadly similar for all types.

The main DRAM variants are:

  • DDR5 (Double Data Rate 5):
    • The latest standard for computer memory
    • Offers higher data rates and improved power efficiency compared to DDR4
    • Used in desktop computers, laptops, and servers
  • LPDDR5X (Low Power Double Data Rate 5X):
    • Designed for mobile devices and other low-power applications
    • Balances performance with energy efficiency
    • Crucial for extending battery life in smartphones and tablets
  • GDDR6X (Graphics Double Data Rate 6X):
    • Specialized for graphics processing units (GPUs)
    • Provides very high bandwidth for demanding graphics applications
    • Used in high-performance gaming cards and professional visualization systems
  • HBM3/E (High Bandwidth Memory 3/E):
    • Designed for applications requiring extreme data transfer rates
    • Particularly important for AI accelerators and high-performance computing
    • Offers the highest bandwidth but is also the most expensive and complex to manufacture

While these variants differ in their peripheral circuits and optimizations, the core memory cell technology remains similar across types.

Source: SemiAnalysis

DRAM Advancements

Recent advancements in DRAM technology have also focused on overcoming scaling challenges and improving performance, with two key innovations standing out: the 4F² cell layout for increased density and Vertical Channel Transistors (VCTs) for enhanced scalability.

The 4F² cell layout represents a significant advancement in DRAM technology. This more compact arrangement of memory cells allows for higher density, potentially doubling the number of cells in the same area compared to previous designs. This improvement is crucial for increasing memory capacity without enlarging chip size.

Second, Vertical Channel Transistors (VCTs) are a novel architecture that enables further scaling of DRAM technology. VCTs allow for better control of electrical characteristics in smaller spaces, which is essential for continuing Moore’s Law in memory development.

These advancements collectively aim to address the growing demand for faster, more efficient, and higher capacity memory solutions across different computing platforms.

 

Main uprising technologies

Outside of DRAM, several companies are working on alternative memory technologies like ReRAM, MRAM, or carbon nanotube-based solutions. The main technologies explored can be divided into two categories:

  1. Short-term solutions:
  • 4F² cell layout: This design allows for a denser arrangement of memory cells, potentially doubling the density compared to current designs.
  • Vertical Channel Transistors (VCT): This new transistor architecture could enable further scaling by improving electrical characteristics in a smaller footprint.
  1. Long-term solutions:
  • Compute-in-memory (CIM) techniques: This approach aims to perform some computational tasks within the memory itself, reducing data movement and potentially improving energy efficiency.
  • New memory types:
    • Ferroelectric RAM (FeRAM): Offers non-volatility and potentially lower power consumption.
    • Magnetic RAM (MRAM): Provides non-volatility and potentially faster write speeds.
  • 3D DRAM architectures: Similar to 3D NAND, this could allow for increased density by stacking memory cells vertically.

 

III – Addressing the Memory Wall Challenge: Vertical Compute’s Innovative Approach

The Memory Wall Challenge

Despite its rapid growth, the memory market faces a critical obstacle: the memory wall. DRAM scaling has slowed considerably, with density increasing only 2x in the last decade compared to the historical 100x per decade. This deceleration creates a bottleneck for computing performance, particularly for AI applications that require vast amounts of memory.

The current limitations are:

  • High Bandwidth Memory (HBM) is effective for AI accelerators but remains expensive and energy-intensive
  • AI applications often rely on cloud-based solutions due to memory limitations, leading to latency issues and potential data privacy concerns
  • Integrated memory (SRAM) is too small, while external memory (DRAM) is too expensive and energy-inefficient

Vertical Compute’s Unique Approach

Vertical Compute distinguishes itself with its innovative vertically integrated memory solution. Unlike other startups focusing on specific technologies, Vertical Compute is developing a revolutionary memory architecture featuring vertical data lanes that would surpass the best attributes of both DRAM and SRAM within a single technology.

Vertical Compute´s technology will unleash the data flow for data-intensive workloads such as generative AI. They are developing a disruptive new technology that, for the first time, integrates vertical data strings/lanes directly on top of the computation units, considerably speeding up data processing. This combo memory and compute solution will be productized as chiplets to Vertical Compute semiconductor customers.

 

Vertical Compute’s solution offers:

  • Performance Gains: 100x performance gains in LLM execution
  • Less power consumption: 80% energy savings by bringing data processing closer to the source
  • Data privacy: with the innovative chiplets, information are processed locally
  • and more, but the company wishes to remain discreet about its technological developments for the time being

Vertical Compute´s technology enables massive time and energy savings thanks to direct, local access between computing cores and data located in a unique vertical high performance and capacity memory. It enables confidential computing and data privacy. This innovative approach could revolutionize the memory landscape, addressing the challenges of capacity, speed, power efficiency, and cost-effectiveness that current technologies struggle to balance. Vertical Compute’s technology promises to combine the density of 3D-NAND with SRAM-like performance, potentially leapfrogging the incremental improvements being pursued by established players. 

 

IV – Team

At Vertical Compute, the strength of the company lies not only in its ambitious vision but also in its founding team. The company was co-founded by Sylvain Dubois and Sébastien Couet, two industry veterans with complementary skills and deep experience in the semiconductor and memory technology sectors.

Sylvain Dubois brings a rare combination of technical and business acumen to Vertical Compute. With a career spanning decades in the semiconductor industry, Sylvain began as a hardware architect before transitioning into business development roles. Notably, he was an early team member and Vice President at CrossBar, a semiconductor startup that successfully raised over $150 million. Most recently, Sylvain worked at Google, where he focused on sourcing advanced technologies, further honing his ability to identify and scale cutting-edge innovations.

Sébastien Couet offers unparalleled expertise in memory technology. Holding a PhD in Physics, Sébastien has dedicated his career to advancing magnetic and oxide heterostructures. At Imec, where he served as Program Director for Magnetics, he led groundbreaking work on MRAM-based memory concepts and other innovative magnetic device technologies. His deep technical knowledge positions Vertical Compute at the forefront of solving critical challenges in memory design. He invented the Verticale Compute technology.

Vertical compute is a spin-off of imec, one of the leading research and innovation centers in Europe, specializing in nanoelectronics and digital technologies. Established in 1984, imec focuses on advancing semiconductor technology and its applications.

We’re delighted to welcome Sylvain and Sébastien to the XAnge family, and in particular to the deeptech vertical. We look forward to supporting them in this ambitious project which, if successful, will have a major impact.

Looking for more content ?

Subscribe to our Newsletter & Follow us on Social Networks

Discover our newsletter