- Tenstorrent Blackhole32 GB · runs 30–34B models at 4-bit
Full specifications
- Processor
- Tenstorrent Blackhole
- Tensix cores
- 120
- Memory
- 32 GB GDDR6, 512 GB/s
- On-chip SRAM
- 180 MB
- Host interface
- PCIe 5.0 x16
- Card-to-card
- 4 × QSFP-DD 800G
- Board power
- 300 W
- Cooling
- Active (fan)
- Software
- Open source: TT-Metalium, TT-Forge
Good for
- Engineers and researchers who want open, programmable AI hardware
- Building multi-card systems over direct 800G links
- Learning and contributing to an open-source AI software stack
Not ideal for
- Anyone who needs CUDA or wants every new model to run on day one
- Running 70B models on a single card
- Gaming or display output
Tenstorrent
Tenstorrent Blackhole p150a
A PCIe AI accelerator built on Tenstorrent's Blackhole chip, with 120 Tensix cores, 32 GB of GDDR6 and four high-speed ports for linking cards together. The software stack is open source, from the low-level kernels up to the model compiler.
£1,169Charged in euros at checkout: €1,399. Import duties and taxes included.In stock. Dispatched in 1 to 3 business days.
- 30-day returns
- Free shipping over €199, charged in euros
- Ships from our UK warehouse
- Import duties and taxes included
- Shipping, returns and warranty
What it can run
Up to 30–34B at 4-bit, 13–14B at 8-bit (approximate)
Based on 32 GB of GPU memory. See every machine by model size
Guides for this product
Local AI · 5 min read
How Much Memory Do You Need to Run a Local LLM?
How much memory a local LLM needs: 7B to 120B+ models at 4-bit and 8-bit, how unified memory differs from VRAM, and which machines hold each size.
Read the guide →Local AI · 5 min read
Mac vs NVIDIA vs Ryzen AI Max+ for Local AI
Which platform fits which work: CUDA on NVIDIA, unified memory on a Mac, or 128 GB on a Ryzen AI Max+ mini PC. Memory, software and trade-offs.
Read the guide →About the Tenstorrent Blackhole p150a
The Tenstorrent Blackhole p150a is a PCIe 5.0 AI accelerator card with 32 GB of GDDR6, 120 Tensix cores, 16 RISC-V cores and four 800G ports for linking cards together. Its software stack is open source from the kernels up. Buy it if you want AI hardware you can program at a low level and scale card by card, and you are prepared to work outside CUDA.
Key features
- Tenstorrent Blackhole processor: 120 Tensix cores and 16 big RISC-V cores
- 32 GB GDDR6 at 16 GT/s, 512 GB/s
- 180 MB on-chip SRAM
- 664 TFLOPS at BLOCKFP8
- 4 × QSFP-DD 800G ports for card-to-card links
- PCIe 5.0 x16, 300 W board power, active cooling
What it runs
On a dedicated card we count about 90% of memory as usable for model weights, which on 32 GB leaves roughly 29 GB. That fits models up to 30B-class at 4-bit and up to 14B at 8-bit. Larger models must be spread over several linked cards. Whether a given model runs at all depends on Tenstorrent's software support for it, so check the model list in their documentation before you buy.
| Model size | 4-bit | 8-bit |
|---|---|---|
| 7–8B | Fits | Fits |
| 13–14B | Fits | Fits |
| 30–34B | Fits | Does not fit |
| 70B | Does not fit | Does not fit |
| 120B+ | Does not fit | Does not fit |
How it differs from a graphics card
Blackhole is built from Tensix cores, each with its own local SRAM, connected by a network-on-chip, plus 16 general-purpose RISC-V cores on the same die. Tenstorrent publishes the low-level programming model, so you can write kernels yourself instead of relying on a closed driver. The card supports FP8, FP16, BFLOAT16, block floating point down to 2 bits, INT8 and TF32 on the Tensix cores.
Scaling across cards
The four QSFP-DD ports on the bracket each carry 800 Gb/s and connect only to other Blackhole cards. Tenstorrent documents topologies for two, four and eight cards. This is the main reason to choose the p150a over the cheaper p100a, which has less memory and no card-to-card ports.
Host system and power
- x86_64 host with a PCIe 5.0 x16 slot and room for a dual-slot card
- At least 64 GB of system memory and 100 GB of storage (2 TB recommended)
- One 12V-2x6 (12+4-pin) power connector and an ATX 3.1 certified power supply
- Up to 300 W per card; about 75 W at idle
- Ubuntu 24.04
The card measures 42 × 270 × 111 mm and cools itself with a fan, so it suits a normal desktop case with good airflow. The passive p150b version is for rack servers only.
Software
TT-Metalium is the low-level programming layer, TT-NN the neural-network operator library above it, and TT-Forge the compiler for models from standard frameworks. All are open source. This is not a CUDA card, and tools written only for CUDA will not run on it.
What to pair it with
A workstation board with a PCIe 5.0 slot, 64 GB or more of RAM and an ATX 3.1 power supply. If you want an easier start on the same software, the Wormhole n300 costs less. If you need broad model support today, a graphics card for local AI is the simpler route.
Good for
- Engineers and researchers who want open, programmable AI hardware
- Building multi-card systems over direct 800G links
- Learning and contributing to an open-source AI software stack
Not ideal for
- Anyone who needs CUDA or wants every new model to run on day one
- Running 70B models on a single card
- Gaming or display output
This is not a CUDA card. Check that the models you need are supported in Tenstorrent's open-source model library before you buy.
Who it's for
Engineers and researchers who want open AI hardware they can program down to the metal and scale card by card.
Full specifications
- Processor
- Tenstorrent Blackhole
- Tensix cores
- 120
- Memory
- 32 GB GDDR6, 512 GB/s
- On-chip SRAM
- 180 MB
- Host interface
- PCIe 5.0 x16
- Card-to-card
- 4 × QSFP-DD 800G
- Board power
- 300 W
- Cooling
- Active (fan)
- Software
- Open source: TT-Metalium, TT-Forge
Reviews
Atomrig customer reviews
No customer reviews yet. Bought this from us? Tell us how it runs.
Related products

NVIDIA
NVIDIA DGX Spark
Runs up to 120B+ at 4-bit
- 128 GB unified
- NVIDIA GB10 Grace Blackwell
Its memory bandwidth (273 GB/s) is lower than a high-end graphics card's, so large dense models generate text more slowly than they would on an RTX 5090. Its strengths are capacity and the CUDA software stack.
£3,929
Tenstorrent
Tenstorrent Wormhole n300s
Runs up to 30–34B at 4-bit
- 24 GB
The 24 GB is split across two chips (12 GB each), so larger models have to be spread across both. Check that the model you want has a working multi-chip version.
£1,169
NVIDIA
NVIDIA GeForce RTX 5090
Runs up to 30–34B at 4-bit
- 32 GB
- NVIDIA Blackwell (GB202)
32 GB runs 30B-class models comfortably at 4-bit, but a 70B model will not fit on one card. For 70B, plan for two cards or a large unified-memory system.
£1,669
PowerColor
PowerColor AMD Radeon AI PRO R9700 32GB
Runs up to 30–34B at 4-bit
- 32 GB
- AMD RDNA 4 (Navi 48)
Check that your tools support ROCm on your operating system before buying. PyTorch and llama.cpp do; some niche tools still assume CUDA.
£1,089