Capabilities · Decision tool

Workload → Hardware Selector

Describe your workload envelope and get the smallest hardware role that can host it — a target requirement, not a measured benchmark.

Choose the smallest and most maintainable platform that satisfies the validated production envelope.

Hyperion does not position itself as a general engineering bench. Technical fluency is used to make better product decisions, challenge assumptions, define evidence and coordinate the critical specialists required by the mandate.

≈ params: <50M · <3B · <14B · <70B · ≥70B

what the system mainly does

response-time requirement

where it physically runs

where data may go

Recommended hardware role

Calculated estimate

NVIDIA Jetson (Thor + Orin)

Fits your envelope

Why this class

  • • On-device perception (cameras) needs a dedicated edge accelerator.
  • • Real-time on-device latency needs a dedicated edge accelerator.
  • • This is the smallest role that satisfies your envelope.
Role
On-robot perception + policy inference at the edge; the workhorse for embedded robotics AI.
Memory
4–128 GB unified (Thor T5000: 128 GB LPDDR5X, ~273 GB/s)
Power
~7–130 W (Orin ~7–60 W; Thor 40–130 W)
Software ecosystem
JetPack, CUDA, TensorRT, ROS 2
Runtimes
TensorRT, CUDA, ONNX Runtime, PyTorch
Known limitations
CUDA-locked; thermal/power budgeting needed for sustained inference. Thor T5000 ≈1035 FP8 TOPS — confirm exact memory bandwidth (≈273–276 GB/s) per the datasheet for a given SKU.
Official source

This recommendation is a calculated estimate over the requirements you entered — a target requirement, not a measurement or a vendor benchmark. Validate against a real workload before committing hardware. Registry specifications are vendor-published and dated; verify them at the official source.

Discuss your hardware envelope
Workload → Hardware Selector — Hyperion