› architecture below the abstraction

hardware

Notes and experiments on Tenstorrent systems—from compute cores and memory movement to multi-chip execution.

01 / understand the hardware

› no black boxes

understand the machine

A reading path for hackers who want to move from the programming model to real kernels, explicit data movement, and the instructions underneath.

Build the mental model: Tensix cores, local SRAM, NoCs, tiles, circular buffers, and reader / compute / writer kernels.

Go from host code to data-movement and compute kernels while learning how a tiled matrix multiply maps onto one Tensix core.

Trace data through unpack, matrix or vector execution, destination registers, and pack—the machinery Metalium kernels drive.

The living low-level reference for Wormhole B0 and Blackhole A0 when an abstraction leaks or a register becomes the shortest path.

02 / build your own

› a compact Tenstorrent workstation

Combine a mini PC, a PCIe dock, and a Blackhole card to build a compact system for developing the software stack.

01 / hardware

Beelink GTi15 Ultra Intel® Core™ Ultra 9 285H 64GB (32GB×2) DDR5 5600MHz + 1TB SSD Beelink Multi-Functional EX Pro Docking Station external PCIe dock + integrated power supply Tenstorrent Blackhole® p150a PCIe accelerator card

02 / setup

01
assemble the system
Seat and power the p150a in the EX Pro dock, then connect the dock to the GTi15 Ultra.
02
install Ubuntu 24.04
Use a clean Ubuntu installation and apply all available system updates.

run the Tenstorrent installer

Install the prerequisites, then run the official setup script.

sudo apt update && sudo apt install -y curl jq
/bin/bash -c "$(curl -fsSL https://tenstorrent.ai/install.sh)"

04
reboot and verify
After restarting, activate the installer-created environment and run tt-smi to confirm that the p150a is detected.