Tenstorrent Wormhole n300 (24GB)

The Tenstorrent Wormhole n300 is not a GPU repurposed for AI — it’s a purpose-built AI accelerator, designed from the ground up by Jim Keller’s Tenstorrent around a fully open-source software stack. Each card carries a pair of Wormhole Tensix processors on a single PCIe board, giving researchers, AI engineers, and hardware enthusiasts a genuinely open alternative to closed-driver GPU stacks — full visibility into the compiler, the kernel implementations, and the hardware itself, with nothing hidden behind a proprietary black box.

Key Features

• Dual Wormhole Tensix Processors — two full ASICs on one card, totaling 128 Tensix cores (64 per processor), each core built around Tenstorrent’s signature array of small RISC-V control processors rather than a fixed-function shader pipeline.
• 24GB GDDR6 memory — running at 12GT/sec for 576GB/s of memory bandwidth, enough to hold meaningfully large quantized models for local inference and experimentation.
• 192MB of on-chip SRAM — 96MB per ASIC, a large, fast local memory pool that Tensix cores can lean on heavily to reduce off-chip memory pressure, a distinctive architectural choice versus traditional GPU cache hierarchies.
• Fully open-source software stack — the entire compiler and runtime (TT-Metalium, TT-NN) is open source on GitHub, giving developers low-level access most GPU vendors never expose, and the ability to hand-tune kernels for their exact workload.
• Native multichip mesh networking — 2x QSFP-DD 400GbE ports plus 2x Warp 100 bridge connectors let you link multiple Wormhole cards into larger mesh configurations, scaling compute and effective memory across cards.
• 1GHz AI clock — delivering up to 466 TFLOPS FP8, 262 TFLOPS BFP8, and 131 TFLOPS FP16 across the card.
• PCIe 4.0 x16 host interface — a standard, drop-in interface for any modern workstation or server motherboard.

A Genuinely Open Platform for AI Research

The Wormhole n300’s appeal is fundamentally different from a GPU’s. Where NVIDIA and AMD cards run AI workloads through layers of proprietary or semi-proprietary driver and compiler stack, Tenstorrent publishes essentially everything — the ISA documentation, the compiler source, the low-level kernel programming model — as open source. For researchers building novel model architectures, kernel engineers optimizing for specific operator patterns, or anyone who’s hit a wall with a closed compute stack that won’t let them see what’s actually happening on-silicon, that openness is the entire value proposition. It’s the same philosophy that made RISC-V attractive at the CPU level, applied to AI accelerator hardware — and Tensix cores are themselves built around arrays of small RISC-V processors rather than a traditional fixed-function GPU pipeline.

Practically, the n300 is well suited to local LLM inference on quantized open-weight models, to multichip AI research where the mesh-networking capability lets you scale beyond a single card’s memory, and to teams who specifically want to develop and profile custom kernels rather than rely entirely on vendor-tuned libraries. It is not a drop-in replacement for a mature CUDA or ROCm stack running mainstream, off-the-shelf inference frameworks — Tenstorrent’s ecosystem, while genuinely capable and rapidly improving, has a smaller library of pre-optimized model support and a steeper learning curve than the incumbent GPU vendors. Buyers should go in expecting to engage with the software stack directly, not simply pip-install a framework and expect instant parity with an NVIDIA card. For the audience it’s built for — open-hardware researchers, RISC-V enthusiasts, and teams who want to build their own optimized inference and training pipelines from first principles — the Wormhole n300 offers a rare combination: real silicon, real memory bandwidth, and nothing hidden.

Specifications

• Processors: 2x Wormhole Tensix Processor (128 Tensix cores total, 64 per ASIC)
• On-Chip SRAM: 192MB (96MB per ASIC)
• Memory: 24GB GDDR6 @ 12GT/sec
• Memory Bandwidth: 576GB/s
• AI Clock: 1GHz
• Compute Performance: 466 TFLOPS FP8 / 262 TFLOPS BFP8 / 131 TFLOPS FP16
• Power: Up to 300W total board power
• Interface: PCIe 4.0 x16
• Networking: 2x QSFP-DD 400GbE + 2x Warp 100 bridge for multichip mesh scaling
• Form Factor: Actively cooled PCIe card
• Software: Fully open-source stack (TT-Metalium, TT-NN)