Specifications

Overview

The Tenstorrent Blackhole p150 is a RISC-V-based AI accelerator built on the Tensix architecture — 120 Tensix cores, each with its own vector engine, SIMD unit and 5 RISC-V baby cores, plus 16 larger RISC-V management cores across the die. The card carries 32 GB of GDDR6 and 180 MB of on-chip SRAM, and runs on a PCIe Gen5 x16 host interface at a 300 W TDP in an actively cooled dual-slot form factor.

Blackhole is designed as an open, scale-out inference and training engine rather than a closed CUDA-alternative: the full software stack (TT-Metalium, TT-NN) is open source, and the p150a variant exposes four passive QSFP-DD 800G ports so cards can be meshed into a rack-scale fabric without a separate switch tier. That makes the p150 a practical route for organisations that want to run large models on non-NVIDIA silicon without being locked to proprietary toolchains.

Note on the specification: Tenstorrent's February 2026 firmware (v19.5.0) reduced the published Tensix core count on p150 cards from 140 to 120 and on-chip SRAM from 210 MB to 180 MB — an effective 1-2% performance reduction on already-shipped units. The figures on this page reflect the current post-firmware specification.

QS Compute can source Blackhole p150 / p150a cards for enterprise and research deployments across APAC. Contact us for card pricing, host-platform compatibility and cluster planning.