Skip to content
MN

//Now · updated 13 October 2026

What I'm doing now.

A now page, in the spirit of nownow.com: the short list of things actually in flight, kept current or deleted.

Updated
13 October 2026
In flight
6
Deliberately not
3

6 active

  1. 01

    Eleven manuscripts in review

    Most of what was rebuilt after the September rejections has gone back in: the thread-scaling paper extended for IEEE TPDS, the decode-roofline paper rebuilt for IEEE IoT-J, and the energy paper re-homed to an Elsevier special issue on energy-aware AI. Now in review: seven at IEEE journals, two at ACM TECS, one at Elsevier SUSCOM, and one at HotMobile 2027. The thermal-margin study was returned on scope and is being re-homed.

    In review
  2. 02

    Artifact evaluation for ATC '26: both reviews delivered

    Both assigned artifact reviews went in inside the window, on 27 September and 2 October. The committee phase runs to mid-October; the badge decisions are the committee’s to make.

    Delivered
  3. 03

    Embedded Vision Summit 2027: talk accepted

    Accepted in September. The next slide drafts are due 4 November and 2 December; the deck is built, checked by a 95-assertion gate against the measured data.

    Preparing
  4. 04

    JOSS reviewing, six assignments deep

    Four reviews completed: nsEVDx, published 6 September as doi:10.21105/joss.11187; Optiland, with the recommendation at the editor; and spherical (VLT/SPHERE high-contrast imaging) and graph-pes (graph-based potential-energy models), each with the full checklist complete. Two are in progress: PySlyde, a pathology preprocessing toolkit, and Battflow, a battery-electrolyte workflow. Six assignments, six different editors.

    Reviewing
  5. 05

    Open-source contributions

    Sixteen pull requests merged into other people's projects: four into ai-dynamo AIPerf, three into NVIDIA garak, a new Arm Learning Path on whether LLM decode is memory-bound on an Arm CPU, six into nsEVDx during its JOSS review and two into pyslyde. llama-roofline is on PyPI, and the quantized-GEMV kernel study is public: twelve CUDA kernels plus Triton, each measured against its own bandwidth roof on five NVIDIA GPUs.

    In motion
  6. 06

    Server-class measurement

    Bringing the Cisco UCS bench up for NUMA and CPU-inference energy work: the same memory-wall thesis, one scale up, on hardware old enough to be honest about.

    Building

Deliberately not doing

  • × Adding a newsletter
  • × Chasing model-of-the-week benchmarks
  • × Writing about tools I have not measured

If this page looks stale, it is. The date at the top is the contract.