Benchmark Library
No tricks, no substitutions. Every benchmark runs the complete network graph with validated accuracy. See the source code, intermediate representations, and full results.
Fully validated with accuracy metrics
CNNs, ViTs, LLMs, Detectors, Segmenters & more
Full end-to-end graphs, no substitutions
Loading benchmarks...
Most NPU accelerators were optimized for ResNet-style networks—simple, repetitive convolutions stacked in sequence. Modern AI looks completely different.
Simple, repetitive 3×3 convolutions. Fixed dataflow patterns.
Multi-head attention, LayerNorm, MatMul. Complex branching graphs.
ResNets to Transformers. Transformers to what's next. AI architectures keep evolving.
Fixed architectures optimized for yesterday's models. Break with every major AI shift.
Fully programmable. Runs any operator, any graph. Ready for whatever comes next.
Our benchmark library grows every quarter with new models across all categories—from classic CNNs to cutting-edge vision transformers and language models.
Every model tested against reference implementations
Access intermediate code and complete network graphs
New models added as the AI landscape evolves
Create a free DevStudio account to access the complete benchmark table with detailed performance metrics, source code, and intermediate representations.
