Blackhole P150 AI Accelerator

Blackhole P150 AI Accelerator | Generated by AI

Home 2026.07

Question: What’s Blackhole@ P150 Series?

Answer:

Blackhole P150 Series is an AI accelerator card from Tenstorrent. It is a competitor to NVIDIA GPUs for AI training/inference, but uses a very different architecture: RISC-V + Tensix AI cores, not CUDA GPUs. (Tenstorrent)

Hardware overview

Blackhole P150:

PCIe Gen 5 AI accelerator card
One Blackhole processor
Up to ~300W power
4 × QSFP-DD high-speed links for multi-card scaling
Designed for AI workloads (training/inference) (Tenstorrent)

The P150 is basically a large AI ASIC + many small RISC-V cores:

Blackhole chip

        RISC-V CPUs
            |
            |
   +----------------+
   | Tensix cores   |
   | AI compute     |
   +----------------+
            |
     SRAM / Memory
            |
       PCIe / Ethernet

Tenstorrent’s design philosophy:

NVIDIA:

CUDA ecosystem
    |
GPU SMs
    |
Tensor Cores

Tenstorrent:

RISC-V control cores
    |
Tensix AI processors
    |
dataflow execution

(Spinics)

Why is it interesting for AI engineers?

The interesting part is software openness.

A Blackhole card can actually run Linux internally because it contains RISC-V cores. The Linux kernel community has added support for Blackhole P100/P150 PCIe cards. (Spinics)

So conceptually:

Normal GPU:

CPU
 |
PCIe
 |
NVIDIA GPU
 |
CUDA kernel


Blackhole:

Host CPU
 |
PCIe
 |
Blackhole
 |
  +-- Linux
  +-- RISC-V cores
  +-- AI accelerator cores

This is closer to a distributed AI computer than a pure GPU.

Performance comparison (rough)

Very roughly:

Card	Architecture	Main ecosystem
NVIDIA H100	GPU + CUDA	dominant
AMD MI300X	GPU + ROCm	open alternative
Tenstorrent Blackhole P150	RISC-V + Tensix	emerging

Blackhole P150 targets:

LLM inference
AI training experiments
large-scale AI clusters

The P150 is not trying to beat H100 immediately on software maturity. Its bet is:

“CUDA is a moat today, but future AI hardware should be programmable and open.”

This is similar to the philosophy behind RISC-V.

For your setup (RTX 4070 + LLM training)

A P150 is interesting because:

Your current stack:

PyTorch
 |
CUDA
 |
RTX 4070

would become:

PyTorch
 |
TT-Metal / Tenstorrent stack
 |
Blackhole

The ecosystem gap is the main issue.

For someone training GPT-2/nanoGPT locally, today:

RTX 4090 / 5090 → easiest
MI300X → serious training
Blackhole → research into future AI architecture

Why people talk about it

Tenstorrent is led by Jim Keller, famous for CPU architecture work (AMD Zen, Apple A-series involvement). The company is trying to build a post-CUDA AI computing platform. (Tenstorrent)

For AI infrastructure builders, Blackhole is interesting because it combines:

open ISA (RISC-V)
AI accelerator
scalable interconnect
Linux-capable compute nodes

Basically: “what if AI clusters were built more like computers, not just GPU farms?”

References:

(Tenstorrent)
(Spinics)

Back Donate