checking fleet…

Mission

Run a mission workload on connected hardware.

Workload

mapping model to chip

CHIP SEES

—

teach an object to begin

–

Ship model

Power and energy for this scene

–Wboard draw

–mJper inference

power ÷ throughput

Accelerator	Power	Scenes/s	mJ/scene
AKD1000checking	–	–	–

Missing same-workload measurements stay blank.

Speed

–fps

single-chip throughput for this workload

RECOGNIZING · ON-CHIP

—

CHIP POWER

–mW

TAUGHT CLASSES

Put an object in the box so it fills it, type what it is, press + TEACH. The AKD1000 learns it live - no retrain, no cloud. Swap in another and teach that. It then recognizes them in real time.

– recorded runs

Run history built for honest hardware comparisons.

Every explicit run records its workload, execution target, timing, throughput, power, energy, and evidence source when those measurements exist. Missing values stay blank.

Workload Execution target

loading run history...

Filtered runs

Explicit hardware executions in this view.

Best throughput

-- fps

Only from rows that reported measured throughput.

Lowest energy

-- mJ

Only comparable within the same workload and input mode.

Comparison rule. Compare hardware only when workload, model, input set, and execution mode match. Dual rows report aggregate throughput across independent partitions. Apex currently runs its vendor COCO reference detector, so it is not an accuracy comparison with the satellite models.

FPGA outputs

Every image here was boxed by LocateAnything 3B on the real Apex FPGA. Timing, power, and satellite accuracy are not claimed on this page.

FPGA harness log

Loading Qwen maritime grounding…

Evidence

Hardware benchmarks

Results are grouped by processor type. Each figure is marked as measured by us, quoted from the vendor, or derived from a stated reference. Anything we have not measured is left blank.

Measured by us on the hardwareVendor specificationReference operating assumptionBlank not measured

Neuromorphic processors

Real Akida silicon, including the on-device learning capability that separates this architecture.

Conventional & programmable accelerators

GPU, FPGA, and AI-accelerator evidence shown separately from neuromorphic results.

On-chip learning. Few-shot learning during inference has been demonstrated on the AKD1000. The current learning model needs 46 neural processors, so it does not fit the 32-NP AKD1500 yet.

AKD1500 energy. The AKD1500 does not report power. Energy per frame uses a 250 mW-per-chip reference divided by measured throughput. This is an assumption, not telemetry.

Loading Overview…

Loading hardware comparison…

Loading Apex telemetry…

Loading orbit view…

Loading AstroAgent…

Loading model zoo…

Loading ocean pass…

Loading Orin Mission Lab…

Loading adaptive hybrid workspace…

Loading evidence lab…