aiWare5 | Next-Gen NPU for Automotive AI Workloads
Smarter means faster: enhanced data-transformation capabilities
Modern neural network architectures — especially Vision Transformers (VT) and State Space Models (SSM) — constantly reshape and reformat data as it flows between layers. These structural transformations can become a bottleneck, leaving compute units idle while waiting for data. aiWare5's new microarchitecture addresses this head-on. It accelerates critical operations like changing activation dimensions, adjusting channel depths, and reordering tensor layouts—all in hardware, with minimal buffer overhead and fewer memory round-trips. The result: MAC units stay busy, and throughput increases significantly for workloads where frequent activation-representation changes are fundamental.
The aiWare Arithmetic Simulator: faster than real time
Companies face critical validation checkpoints throughout the ADAS development cycle: benchmarking quantized networks before deployment, validating software stacks before test-vehicle integration, and certifying production candidates before manufacturing. To enable these activities, organizations routinely deploy hundreds of target hardware units—incurring substantial costs in both bill-of-materials and infrastructure (laboratory space, wiring, cooling, power, and maintenance). The result is expensive and unsatisfying. Hardware-based validation doesn't scale, creates constant bottlenecks, and becomes obsolete with each new chip generation. aiWare-enabled hardware offers an alternative. The aiWare Arithmetic Simulator reproduces every arithmetic operation of the NPU exactly, bit for bit, and runs GPU-accelerated—so you can move all validation activities to cloud or on-premises server farms, leveraging infrastructure you already have.
The benefits: dramatically faster test cycles, significant reductions in capital and operational costs, and virtually unlimited scalability. The infrastructure investment carries forward to your next project; not just the expertise, but the actual compute resources.
LLM support
As the industry standardizes on FP8 for LLM deployment, aiWare5 delivers what most NPUs can't. Customers can add real-time, on-hardware dynamic FP8 scaling to their aiWare configuration — handled natively in hardware, not worked around in software. The result: download an FP8 model from Hugging Face, deploy it directly to aiWare5 hardware, and achieve the published accuracy. No conversion, no fine-tuning, no surprises. It just works.
