Summary
- Daisytuner's €6m seed round was co-led by Elaia and LEA Ventures.
- Its compiler and Daisyflow inference engine are designed to reduce manual optimisation across different processors.
- The company reports Qwen 2.5 running up to 18% faster than vLLM on selected AMD Instinct hardware.
German software developer Daisytuner has raised €6 million to expand technology that automatically optimises artificial intelligence applications for different processors, addressing an engineering constraint that can prevent businesses from making full use of their computing hardware.
The seed round was co-led by Elaia and LEA Ventures, with participation from Amadeus APEX Technology Fund and Seedcamp. The company will use the investment to expand deployment of its Daisyflow inference engine, recruit engineers and develop its compiler technology.
Based in Darmstadt, Daisytuner is developing software that translates applications into processor-specific instructions while automatically identifying opportunities to improve execution.
Daisyflow is already operating in production, according to the company. Initial testing indicates that it can run the Qwen 2.5 model up to 18% faster than the widely used vLLM inference framework on specified AMD Instinct GPUs.
Software determines usable computing performance
AI infrastructure investment is frequently discussed in terms of processor capacity, memory and network bandwidth. However, the performance delivered by that equipment also depends on how effectively software uses the underlying hardware.
Different processors have different memory arrangements, execution methods and programming interfaces. An application designed for one architecture may require additional optimisation before it performs efficiently on another.
That work can involve changing how calculations are scheduled, how information is transferred between memory and processing units, and how operations are divided across available computing resources.
Specialist engineers often carry out these changes manually, using detailed knowledge of the processor and the application being executed.
The process can be expensive and difficult to repeat. Improvements developed for one hardware generation may not transfer directly to another, while changes to an AI model can introduce new performance requirements.
Daisytuner wants to automate more of this engineering work through its Optimizing Compiler Collection, known as DOCC.
The compiler converts code from established frameworks into a representation that can be optimised for different processor architectures.
Its system tests alternative implementations on real hardware, checking whether the proposed changes preserve the intended behaviour of the application.
Verified improvements can then be stored in a performance database and reused when similar workloads are encountered.
This creates an iterative optimisation process in which the system can build on previous results rather than requiring engineers to begin from scratch whenever a new model or hardware configuration is introduced.
Daisyflow connects optimisation to inference
Daisyflow brings the compiler technology into an inference service designed to operate inside customers’ existing infrastructure.
The platform supports language, vision, speech and recommendation models through interfaces intended to work with established AI applications.
Inference systems are responsible for running trained models and producing results in response to user or application requests.
Their performance can influence how many requests a business can process, the time required to generate responses and the amount of infrastructure needed to maintain a particular service level.
More efficient execution may allow an organisation to support greater demand using existing equipment or reduce the computing capacity required for a defined workload.
However, the commercial benefit depends on more than processing speed. Organisations must consider electricity consumption, hardware costs, software integration, maintenance and the operational effort required to support the service.
Response time is also important. An inference engine might process more requests overall without improving the experience of every individual user.
Daisytuner’s initial performance comparison involved Alibaba’s Qwen 2.5 model operating on AMD Instinct GPUs.
The company reports execution up to 18% faster than vLLM under the tested conditions, but has not published sufficient independently verified evidence to establish an equivalent improvement across all models, processors or deployment configurations.
Different workloads can place substantially different demands on memory and processing. Applications using long context windows, large numbers of simultaneous requests or frequent model changes may behave differently from controlled benchmarks.
Customers will therefore need to evaluate Daisyflow using the characteristics of their own applications and infrastructure.
Processor portability expands procurement options
The company’s compiler also addresses the difficulty of moving AI workloads between processor suppliers.
Software developed around one manufacturer’s programming tools may require substantial changes before it can operate efficiently on competing hardware.
These dependencies can influence purchasing decisions even when alternative processors offer competitive theoretical performance or lower acquisition costs.
Daisytuner intends to reduce the engineering effort associated with moving applications between established GPU platforms and newer processor architectures.
Its open-source compiler supports development using frameworks including PyTorch, while the broader platform allows optimisation and performance analysis across different computing environments.
The company has also worked on applications involving emerging hardware, including Tenstorrent accelerators and Q.ANT’s photonic processing technology.
These projects illustrate the range of architectures being explored, although support for particular demonstrations does not establish comprehensive compatibility with every model or commercial application.
Hardware portability could become particularly valuable as businesses evaluate processors developed by new suppliers and attempt to reduce dependence on established software ecosystems.
For European chip developers, usable software tools can influence whether potential customers are able to test their hardware without committing substantial engineering resources.
A compiler capable of reducing that work could broaden the number of processors considered for particular applications.
However, software optimisation cannot eliminate differences in memory capacity, energy consumption or fundamental processing capability.
Hardware selection will continue to depend on the performance, cost, availability and support arrangements of the complete system.
Funding supports broader commercial deployment
Daisytuner was founded in 2024 by Lukas Trümper and Adrian Schmitz, building on research associated with institutions including ETH Zurich, TU Munich, RWTH Aachen and TU Darmstadt.
Its compiler technology reflects research into program transformation and performance optimisation, applying those techniques to the practical demands of increasingly diverse AI processors.
The company says it is working with chip manufacturers, European hyperscalers and enterprise customers.
These relationships may provide opportunities to validate the software against different infrastructure environments, although the company has not disclosed comprehensive commercial deployment figures or independently verified performance results across its supported hardware.
Its development will also require continuing compatibility with new AI models, programming frameworks and processor generations.
Improvements in hardware and software can change which optimisation methods are effective, making regular testing and maintenance necessary even after an application has been deployed successfully.
The €6 million investment will support expansion of Daisyflow and the underlying compiler, but commercial adoption will depend on whether customers can obtain measurable performance improvements without introducing unacceptable operating complexity.
AI inference is becoming an increasingly significant component of enterprise computing expenditure, making the efficiency of existing equipment commercially relevant alongside the development of more powerful processors.
Daisytuner’s early benchmark provides an initial indication of its approach, while broader production results will determine how consistently it can improve performance and reduce the costs of operating AI applications across different hardware platforms.












