SIAgentNews.com
← All news

German startup Daisytuner raises €6M to optimize AI on any chip

Daisytuner is launching a self-learning compiler to help companies run AI models faster on diverse hardware without manual rewriting.

· First reported by eu-startups.com

Claymation: picture for the story 'German startup Daisytuner raises €6M to optimize AI on any chip'
AI-made claymation for this story. No real people or products are shown.
18%Speed increase for Qwen 2.5 on AMD GPUs

The short version

  • Daisytuner raised €6 million in seed funding co-led by Elaia and LEA Ventures.
  • The company launched Daisyflow, a production-ready inference engine for language, vision, and speech models.
  • Their open-source compiler, DOCC, translates PyTorch code into native code for different hardware architectures.
  • The platform reportedly ran Qwen 2.5 up to 18% faster than the industry-standard vLLM on AMD Instinct GPUs.

Why it matters

As AI demand outpaces chip supply, companies need to squeeze every bit of performance out of existing hardware. Daisytuner's automation removes the need for expensive, manual engineering to make models run well on different types of processors.

Breaking the hardware bottleneck

AI companies currently face a major hurdle: compute is a scarce resource, and much of it is locked behind specific software stacks. While many companies rely on NVIDIA, a new wave of accelerators and different chip architectures are emerging to meet growing demand. However, software often prevents these new chips from reaching their full potential.

Daisytuner, a Darmstadt-based startup, aims to solve this by making AI models run efficiently on any hardware without requiring a code rewrite. By automating performance engineering, they want to ensure that every chip runs at full speed the moment it arrives. This approach helps move the industry away from a single-vendor dependency toward a more diverse hardware ecosystem.

The company recently raised €6 million in seed funding to scale its reach. The investment was spearheaded by two European venture capital groups, Elaia and LEA Ventures, with additional backing from Seedcamp, Amadeus APEX Technology Fund, and others. The funding comes as hyperscalers and AI firms shift their focus from simply adding more capacity to optimizing the infrastructure they already own.

How the self-learning compiler works

At the core of Daisytuner's technology is an open-source compiler called DOCC. This tool translates code from popular frameworks, such as PyTorch, into native code specifically for a target chip. This includes GPUs from NVIDIA and AMD, as well as newer, specialized architectures that are just entering the market.

To make this efficient, the company uses a patented optimization platform paired with a self-learning compiler. AI agents work around the clock to test different optimizations on real hardware. Every verified result is then stored in a performance database, which the compiler uses to improve how it handles similar future workloads.

This automated process replaces the traditional method of performance engineering, which has historically been a slow, manual task. In the past, only a few frontier labs had the resources to do this work. By automating the cycle of testing and reusing verified results, Daisytuner allows for faster deployment across a wide variety of hardware configurations.

Launching the Daisyflow engine

Alongside its funding news, Daisytuner launched Daisyflow, a production-ready AI inference engine. Inference is the process of using a trained model to make predictions or generate content. Daisyflow is designed to serve language, vision, speech, and recommendation models using APIs that engineering teams are already familiar with.

One of the primary benefits of Daisyflow is that it can be deployed inside a customer's own infrastructure. This gives companies full control over their data and costs, which is often a priority for enterprises. It allows these organizations to run high-performance models without needing to build massive in-house engineering teams from the very first day.

The platform's capabilities were demonstrated using the Qwen 2.5 model. As reported by eu-startups.com, the system managed to execute the model at speeds up to 18% higher than the standard vLLM when utilizing AMD Instinct GPUs. This performance boost highlights the effectiveness of the underlying compiler and optimization database.

Scaling for a diverse future

Daisytuner was founded in 2024 by Lukas Trümper and Adrian Schmitz. The team brings over a decade of research experience from prestigious institutions like ETH Zurich, TU Munich, and the Lawrence Livermore National Laboratory. They are positioning their technology as the essential software layer for the next era of AI hardware.

The startup plans to use its new capital to grow its engineering team and expand the compiler's support for emerging chip architectures. They are currently working with chip vendors, European hyperscalers, and various enterprises to integrate their tools into production environments. The goal is to make the transition to new hardware as seamless as possible for developers.

As major players like OpenAI and Anthropic begin developing their own custom silicon, the need for hardware-agnostic software becomes even more critical. Daisytuner intends to be the bridge that unlocks performance across every chip on the market. By removing the manual bottleneck of optimization, they aim to let the industry focus on building models rather than wrestling with hardware limitations.

What to watch

  • How quickly Daisytuner can add support for new custom silicon from companies like OpenAI.
  • Whether the 18% speed increase on AMD GPUs becomes a standard benchmark for other model types.

Try it

Check out the Daisytuner Cloud Console to see their live performance database and web IDE for designing custom optimizations.

Sources and links

First reported by eu-startups.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news