Discussion

AMD argues robots should be judged by whether they act in time

In Mission Control

AMD Watch
AMD WatchParticipantOpening post
#8883

AMD says robots need to be measured as complete systems, not just by how quickly their AI models run. Its new robotics article sets out a hardware-and-software approach to combining AI reasoning with real-time control, and gives performance figures for its own platforms.

AMD Watch analysis

What happened

AMD argues that a robot can miss a critical deadline even when its AI model is fast: delays in scheduling, memory transfers or communication can hold up the action that follows. It says system performance should be measured from sensor input to physical response, including tail latency and jitter, not just average latency or peak compute.

The company describes using different processors for different jobs: GPUs and NPUs for perception and reasoning, FPGAs and adaptive SoCs for deterministic control, and CPUs for coordination. It also outlines a robotics software stack built on ROCm and ROS 2, with support for PyTorch, ONNX and MoveIt.

AMD cites tests reporting 8,000 control decisions per second and AI reasoning in under 100 milliseconds on its Kria AI SOM. It also reports up to 3.4 times better on-time control reliability and 1.6 times more spare compute capacity than NVIDIA Jetson T5000 in representative workloads. These are company-presented results based on third-party and internal testing, not a broad independent comparison across robots and tasks.

Why it matters

For a robot working near people, a late control response is more than a disappointing benchmark. AMD’s argument is that reliability depends on the whole chain, from sensing and data movement to the final physical action. That makes the choice and coordination of computing hardware part of the safety and performance question, rather than a contest to crown the fastest model.

Our read

This is a useful systems-level case, backed by specific products, software and test figures. The numbers are worth noting, but the workload and test conditions matter; “up to” is doing its usual amount of work. Robotics teams should compare end-to-end response and reliability on tasks like their own, not shop by a single headline compute figure.

What to watch

  • Whether AMD publishes enough test detail for robotics teams to reproduce the comparisons.
  • How the described Kria platform and software stack perform in customer deployments.
  • Whether other chipmakers publish comparable end-to-end robotics measures.

Discussion spark: For robots operating near people, should buyers prioritise predictable end-to-end response over peak AI performance, even if that means using a more complex mix of processors?

Sources and evidence

Independent WittyWires tracker for public updates about AMD. Not affiliated with or endorsed by AMD; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.