Tenstorrent Wormhole n300 (24GB)

The Tenstorrent Wormhole n300 is 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. It suits researchers, AI engineers and hardware enthusiasts who want full visibility into the compiler, the kernel implementations and the hardware itself.

Who it's for

• Local LLM inference on quantized open-weight models
• Multichip AI research that needs to scale beyond a single card’s memory through mesh networking
• Teams that want to develop and profile custom kernels rather than rely on vendor-tuned libraries
• Open-hardware researchers and RISC-V enthusiasts building their own inference and training pipelines

Key features

• Dual Wormhole Tensix processors: two full ASICs on one card, totalling 128 Tensix cores (64 per processor), each core built around an array of small RISC-V control processors rather than a fixed-function shader pipeline.
• 24GB GDDR6 memory: runs 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 pool that Tensix cores lean on to reduce off-chip memory pressure, unlike a traditional GPU cache hierarchy.
• Fully open-source software stack: the compiler and runtime (TT-Metalium, TT-NN) are open source on GitHub, with ISA documentation and low-level access most GPU vendors never expose, so you can hand-tune kernels for your exact workload.
• Native multichip mesh networking: 2x QSFP-DD 400GbE ports plus 2x Warp 100 bridge connectors link multiple Wormhole cards into larger mesh configurations, scaling compute and effective memory.
• 1GHz AI clock: delivers up to 466 TFLOPS FP8, 262 TFLOPS BFP8 and 131 TFLOPS FP16 across the card.
• PCIe 4.0 x16 host interface: a standard interface for any modern workstation or server motherboard.

Specifications

Good to know

• It is not a drop-in replacement for a mature CUDA or ROCm stack running mainstream, off-the-shelf inference frameworks.
• Tenstorrent’s ecosystem is capable and improving quickly, but it has a smaller library of pre-optimised model support and a steeper learning curve than the incumbent GPU vendors.
• Expect to engage with the software stack directly rather than installing a framework with pip and getting instant parity with an NVIDIA card.

Choose the Wormhole n300 if open hardware and software, multichip scaling and kernel-level control matter more to you than out-of-the-box framework support. If you just want mainstream tools to run, pick a conventional GPU.