Blackhole

Tenstorrent’s Blackhole is a family of PCIe AI accelerator cards designed for developers and researchers who need auditable, open-source hardware for inference and small-scale training. Unlike NVIDIA’s proprietary stack, Blackhole uses a fully open RISC-V instruction set architecture (ISA) and an MIT-licensed compiler stack (TT-Metalium, TT-NN, TT-Forge) that you can read, modify, and submit pull requests to. The cards target workloads where data-movement efficiency and full-stack transparency matter more than raw peak FLOPS — for example, sovereign AI programs, defense and financial services compliance, and hardware research groups. The lineup includes three models: the p100a (active-cooled, no Ethernet, $999), p150a (active-cooled with four passive QSFP-DD 800G ports, $1,399), and p150b (passive-cooled, $1,399). A liquid-cooled desktop workstation, the TT-QuietBox, packs four Blackhole ASICs for $11,999.

Each Blackhole chip integrates 120 Tensix Cores — self-contained compute tiles combining a RISC-V data-movement processor, an 8x8 BF16 matrix engine, and a vector unit — with 1.5 MB of local L1 SRAM per core, yielding 180 MB of on-chip SRAM aggregate. The p100a carries 28 GB of GDDR6 on a 256-bit bus delivering 448 GB/s bandwidth; the p150a and p150b carry 32 GB of GDDR6 at 512 GB/s. Built on a 6nm manufacturing process, the cards draw up to 300W. The p150a’s four QSFP-DD 800G ports enable direct chip-to-chip Ethernet fabric (100 Gbps per link) without InfiniBand or NVLink, scaling from a single card to a 32-chip mesh network. The open-source software stack includes TT-Metalium for kernel development (explicit DMA transfers, tile pipelines) and TT-NN for neural network operators.

Blackhole competes directly with NVIDIA’s RTX 5090 and AMD’s Instinct series, but from a fundamentally different architectural philosophy. NVIDIA’s moat — CUDA, cuDNN, cuBLAS, NVLink, TensorRT — is 20 years deep and closed. Blackhole’s RISC-V ISA and open compiler stack offer full auditability, which is a first-class requirement under the EU AI Act and certain national AI programs. However, the software stack is immature: TT-Metalium kernel development is closer to writing FPGA logic than CUDA, and the ecosystem of pre-optimized models and libraries is thin compared to CUDA’s. In terms of raw memory bandwidth, a single p150a (512 GB/s) trails an H100 SXM5 (3.35 TB/s) by roughly 6x, though the large on-chip SRAM (180 MB) can compensate for small-batch inference where the working set fits locally.

The honest trade-offs are clear. Blackhole offers lower cost per card ($999–$1,399) and full transparency, but at the expense of VRAM capacity (28–32 GB vs. 80 GB on H100), memory bandwidth (448–512 GB/s vs. 3.35 TB/s), and software maturity. The explicit data-movement model makes performance predictable when tuned correctly, but debugging is painful — you must think in terms of tile pipelines and explicit DMA transfers. For teams that need auditable hardware and are willing to invest in compiler-level optimization, Blackhole is a viable alternative. For anyone who just wants to run PyTorch models out of the box with minimal friction, the NVIDIA ecosystem remains the safer bet.

p100a (active-cooled, no Ethernet): $999; p150a (active-cooled, 4x QSFP-DD 800G): $1,399; p150b (passive-cooled): $1,399; QSFP-DD 800G cable: $200; TT-QuietBox (liquid-cooled, 4 ASICs): $11,999.