Running models on your own hardware.
Guides to choosing hardware and running models, with interactive explanations of how the models work.
Harris Oldroyd
I build with AI and explain what I learn.
I run models on my own hardware, test how they perform, and write about how they work. I want to help you understand AI well enough to experiment for yourself. I also write software and research systematic trading.
Testing trading ideas.
Experiments with trading rules, market conditions, and execution costs. I share the results so you can check them yourself.
Software I use in my research.
Tools for collecting data and running experiments locally, so I can repeat a test and check its results.
About me
I'm self-taught.
I learned software and AI through independent study, reading, and building things. I write the explanations I would have found useful while learning.
That sometimes means learning a practical skill too. I've picked up electronics and soldering to investigate hardware questions I couldn't answer from a spec sheet.
Workbench
What I use and how I test.
Run it, measure it, write it up.
I test models, hardware, and runtimes locally. I record the setup and results, and build tools when I need to repeat a test. The write-up explains what I found and the limits of the test.
Threadripper Pro, dual RTX 3090.
- AMD Threadripper PRO 3945WX
- 2 × NVIDIA RTX 3090, 24 GB each
- Debian Server
This is the machine behind the local AI guides. Read the complete workstation build, the RTX 3090 benchmarks, or the multi-GPU motherboard guide.
Apple M5 Ultra, 80-core GPU, 256 GB.
The two NVIDIA DGX Spark orders were cancelled. The current hardware plan is an Apple M5 Ultra Mac Studio with an 80-core GPU and 256 GB of unified memory, on preorder with delivery estimated for October or November.
It has not arrived or been tested here, so no benchmarks or hands-on results from it appear on the site. Read the M5 Ultra article.