Blackhole
AI processor with RISC-V cores and Ethernet scaling
Blackhole is an artificial intelligence accelerator chip designed by Tenstorrent. It uses a mix of large and small RISC-V cores alongside specialised Tensix compute units. The chip relies on standard Ethernet for high-speed connections between multiple devices, aiming to simplify scaling for data centres.
- Uses 16 big RISC-V cores and 752 small cores
- Connects chips via standard Ethernet ports
- Contains 120 Tensix compute cores per chip
- Supports up to 32 GB of GDDR6 memory
- Runs on open-source software stack
What it is
Blackhole is a processor built for artificial intelligence workloads. Tenstorrent designed it to handle heavy computing tasks in servers and workstations. The chip combines general-purpose processing with dedicated math engines. It stands out by using standard Ethernet ports to link many chips together, rather than proprietary cables. This approach allows systems to grow larger without complex custom hardware. The company positions it as a scalable solution for modern AI models.
How it came about
Tenstorrent introduced Blackhole in 2023 as a major update to its earlier Grayskull and Wormhole chips. The design team focused on creating a standalone AI computer that could scale easily. They presented details at the Hot Chips conference in 2024. The goal was to move away from closed ecosystems and offer a more open architecture. This generation marked a shift towards using common networking standards for internal chip communication.
Inside the chip
The chip contains sixteen large RISC-V cores arranged in four clusters. These can run operating systems like Linux. There are also 752 smaller RISC-V cores, known as baby cores, which handle data movement and storage tasks. At the heart are 120 Tensix cores. Each Tensix unit includes five baby cores, two routers, a tile math engine, and a vector math engine. The tile engine works on 32 by 32 data blocks. A network-on-chip connects these units, using static scheduling to move data efficiently. The design keeps data in local SRAM memory whenever possible to reduce delays.
How it was made
Specific manufacturing details such as the fabrication plant or transistor count are not publicly listed in the provided materials. The chip supports up to 32 gigabytes of GDDR6 memory. It operates with a power draw of up to 300 watts in active-cooled forms. The physical implementation comes as PCIe boards, such as the p100a and p150a models, which fit into standard desktop workstation slots.
Where it went
Blackhole processors are used in AI training and inference tasks. They fit into desktop workstations and server racks. The p150a model includes four high-speed ports to link multiple cards. This allows users to pool memory and boost performance across a mesh of chips. Systems like the Galaxy use a 4 by 8 mesh topology to connect 32 chips. The open-source software stack lets developers write custom kernels for specific AI operations.
Versions and family
The p100a variant has 28 GB of memory and no external network ports. The p150a adds 32 GB of memory and four QSFP-DD 800G ports for linking cards. Both use the same underlying Blackhole Tensix Processor with 120 cores. They replace the earlier Wormhole n150 and n300 boards. The Wormhole series remains available but is considered the previous generation technology.
Why it matters
Blackhole promotes the use of Ethernet for scaling AI hardware. This choice leverages industry-standard networking gear instead of proprietary links. It supports a fully open-source software model, giving developers direct access to the hardware. This approach aims to lower barriers for engineers building custom AI solutions. It challenges the dominance of traditional graphics processors in the AI market by offering a different architectural path.
Questions people ask about Blackhole
What makes Blackhole different from other AI chips?
It uses standard Ethernet for connecting multiple chips instead of proprietary links. It also features a mix of large and small RISC-V cores alongside specialized Tensix units, all supported by open-source software.
How many cores does the Blackhole chip have?
The chip has 16 large RISC-V cores that can run Linux, 752 smaller baby RISC-V cores for data tasks, and 120 Tensix cores dedicated to math and compute operations.
Can I connect multiple Blackhole cards together?
Yes, the p150a model includes four QSFP-DD 800G ports. These allow you to link multiple cards to pool memory and improve performance in a mesh topology.
What software runs on Blackhole?
It supports a fully open-source software stack. Developers can use TT-Metalium for low-level access or TT-Forge for higher-level development, granting full access to the hardware.
Good reading and sources
There is no English Wikipedia article on Blackhole yet, so the facts on this page come from the sources listed above: tenstorrent.com/hardware/blackhole, servethehome.com/tenstorrent-blackhole-and-metalium-for-standalone-ai-. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.