Our testing infrastructure
Multiple vendors, multiple architectures — side by side, ready for independent benchmarking. All available for remote testing.
Our lab is constantly expanding. Have a specific device you'd like us to test? Let us know.
From data center racks to pocket-sized edge chips
Three tiers of AI hardware, covering every deployment scenario.
Data center inference
Run frontier models locally. Multi-GPU and multi-chip systems for large-scale inference, from single accelerator cards to liquid-cooled multi-chip platforms. The hardware that powers production AI.
NVIDIA The industry standard
CUDA-based GPUs with the broadest software ecosystem. The default choice for most AI workloads — and the benchmark everything else is measured against.
AMD The GPU alternative
Instinct accelerators with growing ROCm software support. A competitive option for inference workloads, often at lower cost than NVIDIA equivalents.
Tenstorrent The open-source RISC-V path
RISC-V-based AI accelerators with a fully open-source software stack. Up to 5x lower TCO than NVIDIA for inference workloads. From single PCIe cards to liquid-cooled multi-chip systems.
Q.ANT Photonic computing from Germany
Next-generation photonic AI processors that compute with light instead of electrons. Up to 30x energy efficiency versus traditional silicon. Made in Germany, funded by BMBF.
Desktop AI & local inference
AI on your desk. Single-card accelerators, desktop AI supercomputers, and unified-memory workstations that let you run large models without a server room. The answer to "can I run this model locally?"
NVIDIA Desktop AI supercomputer
The DGX Spark brings Grace Blackwell to a desktop form factor with 128 GB unified memory — powerful enough for most production models without a server room.
Tenstorrent Affordable single-card AI
A single Blackhole PCIe card delivers 664 TFLOPS for under €1,300 — plug it into any workstation and start running models. The QuietBox packs four chips in a quiet, liquid-cooled desktop.
Apple Unified memory for large models
Mac Studio with M-series Ultra offers up to 192 GB unified memory — enough to run 70B models at FP16 or 120B+ at INT4. The best price-per-GB for local inference in the Apple ecosystem.
On-device inference without the cloud
AI at the edge — offline, sovereign, power-efficient. Run vision models and small LLMs on devices that draw less power than a phone charger. Perfect for manufacturing floors, retail, vehicles, and anywhere cloud connectivity isn't guaranteed.
Hailo The edge AI leader
Dataflow architecture designed for maximum power efficiency. The Hailo-10 brings LLM inference to edge devices at just 2.5W. Deep Raspberry Pi integration makes prototyping fast.
Axelera European RISC-V edge AI
Dutch-designed edge processors using Digital In-Memory Computing on RISC-V. The Metis delivers 214 TOPS at extreme power efficiency. EU-funded, European supply chain.
DEEPX Ultra-low-power embedded AI
South Korean edge AI chips optimized for the absolute lowest power envelope. The DX-M1 delivers 25 TOPS at just 1-5W in a tiny M.2 form factor.
NVIDIA CUDA at the edge
Jetson brings the CUDA ecosystem to edge devices. Familiar tools and frameworks for developers already in the NVIDIA ecosystem, with GPU-accelerated inference on-device.
Real numbers, not marketing claims
Every benchmark report includes the metrics that matter for production deployment.
Throughput
Tokens/second for LLMs, frames/second for vision. Under real-world concurrent load, not synthetic peaks.
Latency
Time to first token, P50/P95/P99 latency. The numbers that determine whether your users wait or don't.
Power
Per-device watt measurements. Critical for edge deployment and data center TCO calculations.
Cost-per-inference
Hardware purchase price amortized to cost per million tokens. The number your CFO cares about.
Compatibility
Did the model need conversion? Quantization? What broke? Honest notes on real-world readiness.
Methodology
MLPerf-aligned, fully documented, open-source benchmark scripts. Reproducible by anyone.
Want to test your model on our hardware?
Book a consultation and we'll design a benchmark plan for your specific workload.