01 / THE PLATFORMA CONNECTED SYSTEM

Three stages.
One system.

PreprocessingInferencePostprocessing

Three bodies, one shared orbit. Arrows show acceleration.

Designing all stages together makes better use of the whole system.

In many applications, preprocessing, inference and postprocessing are developed as independent parts. A stage typically either limits the other stages or has capabilities that the rest of the system leaves unused.

With spacetime AI, they work in unison. Each stage is designed and adapted to help the others perform better.

Preprocessing assembles more densely informative inputs: more useful information in less data, making better use of processing time and memory. Models are adapted to learn from these inputs and produce richer outputs, giving postprocessing the information it needs to deliver better results.

Like space and time, these stages belong together: a single, unified system.

Preprocessing

Assemble densely informative inputs: less data, more useful information. Prepare what the model needs with the final postprocessing results in mind.

Inference

Adapt models to use those richer inputs and return more useful information for postprocessing.

III

Postprocessing

Turn richer model outputs into useful results for your application, using information that preprocessing and inference were designed to provide.

HARDWARE CHOSEN FOR THE SOFTWARE

More performance doesn’t
have to mean more power.

Start with what the software needs to do. Choose hardware built for that kind of work, then adapt the software to make full use of it. The hardware follows the needs of the application.

For our vision pipelines, that means selecting the right combination of CPU, GPU, AI accelerator, video engines and unified memory. The models and processing stages are designed around how these parts work together, reducing unnecessary computation and data copying.

Demanding vision workloads do not automatically need a large, power-hungry machine. For suitable workloads, our matched hardware and software outperform a much larger workstation while using far less power.

This approach supports both fast vision models, including CNNs, and newer and experimental transformer models.

Performance, power use and model support depend on the workload and device.

✳Models + pipelineENGINEERED TOGETHER

CPUGPUNPUVIDEO ENGINESHARED MEMORY ARCHITECTURE

Hardware we usually work with

The AI universe is huge and varied. Each of these platforms has its place: different tasks call for different strengths. We choose the hardware to suit the application.

NEXT / DEVELOPERSDifferent devices. The same interface.