Vulkan Performance on NVIDIA hardware
NVIDIA's llama.cpp benchmarks show Vulkan matching or exceeding CUDA on selected workloads using its Cooperative Matrix 2 extension.
Our mission is to help teams carry their AI software across hardware platforms.
More choice, more integration
Robots and edge devices have more GPUs and accelerators to choose from, each with its own software stack. A hardware change means integrating and validating the pipeline again. Physical AI needs a common software layer so teams can reuse more of their work.
A common layer on Vulkan
We build our AI inference runtime on Vulkan, an open standard for GPUs across vendors. It combines vAi, our inference engine, with vStream, our pipeline connecting sensor input, processing and output.
Extending beyond GPUs
Qualcomm's vendor extension defines neural processing engines as targets for Vulkan data graphs. Support depends on the device and driver.
Contact
- General — contact@vaistrata.com
- Security — security@vaistrata.com