VeizikNVIDIA, Apple Silicon and edge builds are available now.Release notes →
veiziknative inference engine
For developers Sign inDiscuss deployment
VEIZIK for business

Keep control responsive while AI runs.

Evaluate vision-language-action inference and CPU control scheduling on the same device, with separate measurements for latency, fidelity and deadlines.

Evaluate your control workloadSee the technology →

Technical demonstrations and measured records only. No customer deployment is implied.

GPU lane — VLA inferenceA vision-language-action policy forward pass. Evaluated on its own timeline: latency and output fidelity.
Shared deviceOne board, one power mode, pinned clocks. Other GPU workloads declared, not assumed idle.
CPU lane — control loopPeriodic loop on an isolated core, SCHED_FIFO, memory locked. Evaluated by response time and deadline misses.
Reported separatelyModel forward · control-loop response · sensor-to-action. Never merged into one number.
Three conditions

Latency, fidelity and deadlines, kept apart.

Each condition has its own boundary and its own result. None stands in for another.

Latency

Two different questions. How long the model forward takes, and how quickly the control loop produces an action once its release instant arrives. They are measured at different boundaries and shown in different boxes.

Fidelity

Each action is compared with the full-precision reference output by cosine similarity. That measures how closely the policy output is preserved. It is not task success and is never presented as such.

Deadlines

Missed releases are counted over a stated number of cycles, with the scheduler, core isolation and kernel recorded. A cadence such as 50 Hz or 1 kHz is a cadence we measured, not a current product guarantee.

Results

Two number boxes, never merged.

Measured records where they exist. Where a configuration has not been measured, the box says so.

Box 1

VLA model forward

Not measured for this configuration

The published robotics measurements on this site are for compact control policies, not a vision-language-action model. We do not substitute one for the other.

Box 2

Control-loop response

Engine latency in microseconds: from the expected release instant to action-ready, taken inside the runtime with the monotonic clock. Compact control policy on the CPU loop.

Scenariop50 µsp99.9 µsmax µsjitter p99.9 µsDeadline misses
50 Hz control (idle)21273100
50 Hz + full CPU saturation224145≤30
1 kHz torque-rate control21242900 / 20,000

Jetson AGX Orin 64 GB, 12× Arm Cortex-A78AE, L4T R36.4.3, Linux 5.15 stock (PREEMPT, not real-time), clocks pinned, one CPU core isolated, SCHED_FIFO with locked memory and absolute-time periodic release.

Sensor-to-action, end to end

Not measured for this configuration

System latency from sensor callback to actuator is a separate, larger boundary. It is never conflated with engine latency, and no full measurement is published here.

Sustained run

0 / 180,000 missed deadlines at 50 Hz and 0 / 3,600,000 at 1 kHz over one hour each. Response p99.9 / max: 26 / 45 µs at 50 Hz, 21 / 36 µs at 1 kHz. Average / peak CPU power at 50 Hz: 2.4 / 3.2 W.

A raw-timestamp verifier outside the engine recomputed the counts. Core isolation was required: without it, rare kernel-interference events produced occasional multi-millisecond stalls. This describes one measured run, not a guarantee.

Action fidelity

Cosine similarity to the full-precision reference across four compact control policies: 0.99978 to 0.99999.

This is output similarity, not task success.

Task success

Not measured for this configuration

No task-suite result is published. Any future success figure will show numerator and denominator with the checkpoint and task split.

CPU control with GPU VLA running together

Not measured for this configuration

The measurements above are control-loop runs. A combined run with a VLA policy active on the GPU is the configuration an evaluation would measure.

Integration boundary

What is measured, and where it stops.

What the runtime covers

A self-contained runtime generated by VEIZIK’s own compiler, with a per-device activation. It runs the policy and reports latency and action fidelity from the command line.

What stays with you

Sensors, actuators, the robot’s own safety systems, the real-time kernel choice and the system-level integration. Whole-robot safety is outside what VEIZIK measures or claims.

Stated limits

Figures come from one board class (Jetson AGX Orin) on a stock kernel. They are not a head-to-head comparison with results published on other devices.

Same board, other workloads

One device runs the control loop and the generative stack.

Robotics evaluations weigh what else shares the GPU. The Jetson AGX Orin that schedules your control loop also runs the full image stack natively — these were drawn on it.

A rain-soaked cyberpunk megacity at night from a rooftop, holographic billboards in magenta and cyan, neon haze, light trails — AI-generated with FLUX.1-schnell on veizik, rendered on an NVIDIA Jetson AGX Orin.
FLUX.1-schnellNVIDIA Jetson AGX Orin
Extreme macro of a dew-covered jumping spider on a leaf at dawn, iridescent eyes, water droplets refracting light — AI-generated with FLUX.1-schnell on veizik, rendered on an NVIDIA Jetson AGX Orin.
FLUX.1-schnellNVIDIA Jetson AGX Orin
A delicate botanical watercolour of a flowering magnolia branch, antique herbarium-plate style, blush-pink petals on cream paper — AI-generated with FLUX.1-schnell on veizik, rendered on an NVIDIA Jetson AGX Orin.
FLUX.1-schnellNVIDIA Jetson AGX Orin
An Art Nouveau decorative poster in the style of Alphonse Mucha — an elegant woman wreathed in white lilies, ornate golden halo, floral border — AI-generated with FLUX.1-schnell on veizik, rendered on an NVIDIA Jetson AGX Orin.
FLUX.1-schnellNVIDIA Jetson AGX Orin
A charming isometric low-poly 3D render of a tiny floating island with a windmill, pastel cube trees and a pond — AI-generated with FLUX.1-schnell on veizik, rendered on an NVIDIA Jetson AGX Orin.
FLUX.1-schnellNVIDIA Jetson AGX Orin

Shown as what the engine draws on the board — output, not a benchmark; no latency or throughput figure is attached to these images. Each is a FLUX.1-schnell image rendered natively by Veizik on a Jetson AGX Orin, the same class of board that schedules a control loop.

Evaluation

Bring the workload. We measure the fit.

A technical evaluation starts from the inputs below.

Evaluate your control workloadSee the technology →