Research

Research that ships.

One question: how do you make AI fast and cheap enough to live in the physical world? Here is everything we have run, with what worked, what didn’t, and the numbers behind both.

  • Efficient inference
  • Real-time runtimes
  • Energy-aware compute
  • Robust perception

  1. Experiment E004

    Hybrid serving of large mixture-of-experts models

    Experts held in system memory, everything else on one GPU. Can a memory-rich server serve large MoE models at a lower cost per token than GPU infrastructure sized to fit them?

    In progress

  2. Experiment E003

    Why AMX bf16 matrix multiply scales poorly across cores

    E001 showed Intel AMX throughput falling as threads were added: 13.5 TFLOP/s at 8 threads, 5.0 at 16. Up to 24 threads we have ruled out measurement noise, kernel switching and power limits. Thread placement, huge pages and framework dispatch are still untested.

    Partial result

  3. Experiment E002

    CPU-only language-model serving

    Serving dense models up to 8B parameters on CPUs alone cost 7 to 20 times more per token than a fully loaded H100. We rejected the hypothesis and moved on to hybrid serving.

    Result

  4. Experiment E001

    CPU vs GPU: dense compute and small-model training

    Dense compute cost 8 to 67 times more per operation on CPU than on GPU. A clear negative result, and the baseline for everything after it.

    Result

Every number traces to raw measurements and the scripts that produced them. Papers, benchmarks and open-source tools will be listed here as they are released.

Build physical AI that works everywhere.

Get early access Talk to the lab