Artificial by Design provides a GPU/CPU hypervisor that acts as a compatibility layer, intercepting CUDA API calls and translating them so AI workloads written for NVIDIA GPUs can run unchanged on AMD, Intel, or custom AI accelerators. The software‑based runtime engine, optionally accelerated by CPUs or hardware offload, delivers performance that can exceed native CUDA while allowing enterprises to diversify hardware and reduce infrastructure costs.
Funding
Funding not disclosed
Founders
Product
Problem
AI applications are tightly coupled to NVIDIA’s CUDA ecosystem, making it difficult to run workloads on alternative GPU or custom accelerator hardware. This lock‑in forces organizations to invest in specific hardware and limits flexibility in heterogeneous AI infrastructure.
Solution
Artificial by Design offers a GPU/CPU hypervisor that acts as a compatibility layer for AI workloads. By intercepting CUDA APIs, the ABD Runtime Engine translates and maps kernel calls to a range of accelerators—including NVIDIA, AMD, Intel GPUs and custom AI ASICs—without requiring any changes to the original PyTorch, TensorFlow, vLLM, or ONNX code. The translation is performed in software, accelerated by standard enterprise CPUs, and can be further boosted with optional hardware acceleration, delivering performance that can exceed native CUDA execution. This approach enables organizations to repurpose existing CPU servers, diversify hardware vendors, and reduce infrastructure costs while maintaining full functionality of their AI models.
Target Audience
Primary customers are enterprise data centers, cloud service providers, and large‑scale AI teams that need to run CUDA‑based workloads on diverse or existing accelerator hardware.
Features
- Compatibility layer intercepts CUDA API calls and provides zero‑code‑change migration for existing AI applications
- Runtime engine performs kernel translation and execution mapping across heterogeneous accelerators
- Supports major AI frameworks (PyTorch, TensorFlow, vLLM, ONNX) out of the box
- Optional CPU‑based acceleration and hardware‑offload paths to improve performance beyond native CUDA
- Works with NVIDIA, AMD, Intel GPUs and custom AI ASICs, enabling true multi‑vendor AI infrastructure
- Provides a hypervisor abstraction that isolates workloads from underlying hardware specifics