← Back to homepage
System highlight

The COGG9 work is tested on a local AI workstation.

This is the machine I use for local AI experiments, model direction tests, compression/signal work, and media generation trials. It keeps the research grounded because the hardware has limits.

local machine / real limits / test bench
01CPU

AMD Ryzen 9 3900X, 12 cores and 24 threads.

02Memory

64 GB system RAM for local development and test runs.

03GPU

Two Intel Arc Pro B70 GPUs, each reporting about 30 GB VRAM.

04OS

Ubuntu 24.04 LTS on a Linux 6.17 kernel.

Why show the machine?

Because the hardware matters.

A lot of AI talk floats away from the actual machine doing the work. I do not want that here. COGG9 is about local-first AI research, so the test box is part of the story.

This does not make the work better by itself. It just makes the pressure visible: memory, compute, drivers, storage, render time, failures, and the weird limits that show up when the work is not just an idea.

Processor

AMD Ryzen 9 3900X, 12 cores / 24 threads.

System memory

64 GB RAM. Enough to do real local work, not enough to pretend memory is infinite.

Graphics / AI accelerators

Two Intel Arc Pro B70 GPUs. OpenCL reports each card with about 30.3 GiB global memory.

Operating system

Ubuntu 24.04 LTS with Linux kernel 6.17.

Motherboard class

ASUS ROG Crosshair VIII Hero Wi-Fi platform.

Storage

Local SSD/NVMe and HDD storage for experiments, generated media, checkpoints, and working artifacts.

What this system is used for

This is the local machine behind a lot of the current COGG9 work: AI harness testing, model experiments, compression/signal notes, video generation trials, local tooling, and general build work.

It is not a polished lab cluster. It is a working machine. That is useful. The rough edges show what has to be handled if the research is going to stay close to normal hardware.

Why limits matter

Local hardware changes the questions.

Memory pressure, GPU support, driver behavior, storage, render time, and recovery all become part of the work. That is one reason COGG9 Agent, COGG9, and WaveCodecLLM fit together for me: the model, harness, and signal path all have to respect the machine they run on.

The workstation is not the whole story. It is the current test bench.

What this page is not

This is not a benchmark claim, a hardware endorsement, or a promise that every experiment runs cleanly. It is just a simple public note about the system being used right now.

Some details stay private: exact file paths, internal scripts, protected mechanics, credentials, and raw lab artifacts.