PantheonSim / GPUs / Tenstorrent machines
Simulated Tenstorrent machines: Blackhole and Wormhole
PantheonSim, the GPU simulator for CUDA and ROCm, also offers 4 Tenstorrent machines in the playground: Blackhole and Wormhole, with one chip or two. They have no profile like the NVIDIA and AMD cards do. They run on Tenstorrent’s own simulator, ttsim, which is a functional simulation with no performance data.
The machines
| Machine | Architecture | Chips | Tensix cores per chip | DRAM | L1 per Tensix |
|---|---|---|---|---|---|
| Tenstorrent Blackhole | Blackhole | 1 | 130 (13 x 10) | 31.9 GiB | 1536 KiB |
| Tenstorrent Blackhole, two chips | Blackhole | 2 | 110 (11 x 10) | 2 x 31.9 GiB | 1536 KiB |
| Tenstorrent Wormhole | Wormhole | 1 | 72 (8 x 9) | 12 GiB | 1464 KiB |
| Tenstorrent Wormhole, two chips | Wormhole | 2 | 56 (8 x 7) | 2 x 12 GiB | 1464 KiB |
A two-chip machine is a different machine, not a count of two: it has its own core grid, and the second chip is reached over Ethernet, so a first program takes longer to start. The figures come from the machine catalog the playground itself reads, so this table is what a session sees.
How they work
A session starts a login shell on Ubuntu 24.04 with tt-metal installed, and ttsim stands in for the chip. Programs written for
tt-metal run unmodified. PantheonSim’s own profiles, and commands such as vgpu run --gpu, do not apply to
these machines. They are chosen in the playground, one machine per session.
The simulation is functional. It checks what a program computes and that it uses the API correctly, and it says nothing about how fast a real chip would run it. The Tenstorrent simulator guide walks through a first program.
Also: NVIDIA GPUs, AMD GPUs, AWS Neuron machines, or all of them.
Questions
Which Tenstorrent chips can I try?
Blackhole and Wormhole, each as a one-chip and a two-chip machine, in the playground.
Can the simulator tell me how fast my kernel is?
No. ttsim is a functional simulator and reports no performance figures.