Tenstorrent Blackhole p150a (32GB)

The Tenstorrent Blackhole p150a is Tenstorrent’s newest and most powerful single-card AI accelerator, an actively-cooled PCIe card built around the Blackhole architecture — the successor to Wormhole — and aimed at researchers, kernel engineers, and AI developers who want serious on-card compute and a fully open software stack in one desktop- or workstation-ready package. Where Wormhole proved the open-hardware approach works, Blackhole scales it up, packing more Tensix cores, more on-chip SRAM, and faster memory into a single-ASIC design.

Key Features

• 120 Tensix Cores — Tenstorrent’s core compute unit, each built around a cluster of small RISC-V processors rather than a traditional GPU shader pipeline, giving fine-grained programmability over how work is scheduled and executed.
• 16 “big” RISC-V cores — dedicated general-purpose RISC-V processors on-die for control, orchestration, and running host-side logic close to the compute, reducing round-trips to the CPU.
• 180MB of on-chip SRAM — roughly 1.5MB per Tensix core, a large local memory pool that keeps data close to the compute units and reduces pressure on off-chip DRAM bandwidth.
• 32GB GDDR6 on a 256-bit bus — running at 16GT/sec for 512GB/s of memory bandwidth, enough headroom for meaningfully large local models and datasets.
• Up to 1.35GHz AI clock — delivering 664 TeraFLOPS at BLOCKFP8 precision, a substantial generational step up from Wormhole.
• 4x QSFP-DD 800G ports — passive high-speed networking that lets multiple Blackhole cards link together to pool memory and scale compute across a multi-card system.
• PCIe 5.0 x16 host interface — the latest-generation host interconnect for maximum data throughput between card and system.
• Fully open-source software stack — the same open compiler and runtime approach (TT-Metalium, TT-NN) that defines Tenstorrent’s entire product line, with nothing hidden behind proprietary drivers.

Tenstorrent’s Most Capable Open AI Accelerator

Blackhole represents a real architectural step forward from Wormhole, and the p150a is the card that showcases it best on a single-slot budget: more Tensix cores, nearly double the SRAM density per core, faster GDDR6 memory, and a move to PCIe 5.0 and 800G networking that positions it for genuinely serious multi-card scale-out. For AI research teams building or fine-tuning custom model architectures, kernel engineers who want to hand-optimize compute graphs at a level GPU vendors simply don’t expose, and RISC-V-focused developers who want general-purpose cores sitting directly alongside the AI compute fabric, the p150a is currently Tenstorrent’s flagship answer.

The 16 on-die “big” RISC-V cores are a distinctive feature worth calling out: they let orchestration, data preprocessing, and control-flow logic run directly on the card rather than round-tripping to the host CPU for every step, which matters for latency-sensitive or highly custom inference and training pipelines. Combined with 180MB of SRAM, the architecture is built to minimize the memory bottlenecks that constrain many accelerator designs, and the 800G QSFP-DD ports mean a multi-card Blackhole cluster can scale bandwidth well beyond what standard PCIe-only multi-GPU setups typically achieve.

As with every Tenstorrent product, the honest trade-off is software maturity versus openness. TT-Metalium and TT-NN are genuinely open, actively developed, and give you full visibility into how your workload maps to silicon — but the ecosystem of pre-optimized model support is smaller than CUDA’s or even ROCm’s, and getting the most out of the hardware generally means engaging directly with Tenstorrent’s programming model rather than expecting instant, zero-effort parity with a mainstream GPU framework. For teams who value that openness and are willing to invest in the stack, the p150a offers real, differentiated compute — and one of the only actively-cooled, single-card AI accelerators on the market where you can read and modify every layer of the software that runs on it.

Specifications

• Tensix Cores: 120
• RISC-V Cores: 16 (“big” RISC-V, on-die)
• On-Chip SRAM: 180MB (~1.5MB per Tensix core)
• Memory: 32GB GDDR6, 256-bit bus @ 16GT/sec
• Memory Bandwidth: 512GB/s
• AI Clock: Up to 1.35GHz
• Compute Performance: 664 TeraFLOPS (BLOCKFP8)
• Power: 300W total board power, 1x 12+4-pin (12V-2×6) connector
• Interface: PCIe 5.0 x16
• Networking: 4x QSFP-DD 800G (passive) for multi-card scale-out
• Dimensions: 42mm x 270mm x 111mm, actively cooled
• Software: Fully open-source stack (TT-Metalium, TT-NN)