changh95

/

superpoint-p150

Like 1

This adds a p100 serve profile so the bundle can be served on a Tenstorrent p100 (one Blackhole chip, 7 DRAM banks instead of the p150's 8).

I ran every single-chip p150 community bundle on a p100a (2026-10-15/16), and this one worked with its existing p150 settings unchanged, so the new profile is a copy of the tested profile with hardware: p100.

What changes

• tt_kernel_manifest.json: a new p100 profile. Everything except name, hardware and description is identical to the profile that was tested, and mesh_device stays P150 because that is what ran. The default profile is unchanged, so nothing changes for existing users.
• README.md: the p100 tag (so tt lists the bundle as p100-compatible).

No code or image changes; the bundle does not need to be rebuilt.

Verified on the p100a: 500 keypoints with 256-d descriptors, all finite, 5.4 ms on device.

Notes

• Your card mentions dispatch on the Ethernet cores. A p100 has no active Ethernet, so its latency may differ from the p150 numbers; the README note says they were not re-measured.

To try it once merged: tt-model serve changh95/superpoint-p150 --profile p100 (or tt serve changh95/superpoint-p150 --profile p100).

🤖 Generated with Claude Code