{"title":"Tenstorrent","description":null,"products":[{"product_id":"tenstorrent-wormhole-n300-24gb","title":"Tenstorrent Wormhole n300 (24GB)","description":"\u003cp\u003eThe 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.\u003c\/p\u003e\n\u003ch2\u003eWho it's for\u003c\/h2\u003e\n\u003cul\u003e\n\u003cli\u003eLocal LLM inference on quantized open-weight models\u003c\/li\u003e\n\u003cli\u003eMultichip AI research that needs to scale beyond a single card’s memory through mesh networking\u003c\/li\u003e\n\u003cli\u003eTeams that want to develop and profile custom kernels rather than rely on vendor-tuned libraries\u003c\/li\u003e\n\u003cli\u003eOpen-hardware researchers and RISC-V enthusiasts building their own inference and training pipelines\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003ch2\u003eKey features\u003c\/h2\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cstrong\u003eDual Wormhole Tensix processors:\u003c\/strong\u003e 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.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e24GB GDDR6 memory:\u003c\/strong\u003e runs at 12GT\/sec for 576GB\/s of memory bandwidth, enough to hold meaningfully large quantized models for local inference and experimentation.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e192MB of on-chip SRAM:\u003c\/strong\u003e 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.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eFully open-source software stack:\u003c\/strong\u003e 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.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eNative multichip mesh networking:\u003c\/strong\u003e 2x QSFP-DD 400GbE ports plus 2x Warp 100 bridge connectors link multiple Wormhole cards into larger mesh configurations, scaling compute and effective memory.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e1GHz AI clock:\u003c\/strong\u003e delivers up to 466 TFLOPS FP8, 262 TFLOPS BFP8 and 131 TFLOPS FP16 across the card.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003ePCIe 4.0 x16 host interface:\u003c\/strong\u003e a standard interface for any modern workstation or server motherboard.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003ch2\u003eSpecifications\u003c\/h2\u003e\n\u003ctable\u003e\n\u003cthead\u003e\u003ctr\u003e\n\u003cth\u003eSpec\u003c\/th\u003e\n\u003cth\u003eDetail\u003c\/th\u003e\n\u003c\/tr\u003e\u003c\/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd\u003eProcessors\u003c\/td\u003e\n\u003ctd\u003e2x Wormhole Tensix Processor (128 Tensix cores total, 64 per ASIC)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eOn-chip SRAM\u003c\/td\u003e\n\u003ctd\u003e192MB (96MB per ASIC)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eMemory\u003c\/td\u003e\n\u003ctd\u003e24GB GDDR6 @ 12GT\/sec\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eMemory bandwidth\u003c\/td\u003e\n\u003ctd\u003e576GB\/s\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eAI clock\u003c\/td\u003e\n\u003ctd\u003e1GHz\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eCompute performance\u003c\/td\u003e\n\u003ctd\u003e466 TFLOPS FP8 \/ 262 TFLOPS BFP8 \/ 131 TFLOPS FP16\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003ePower\u003c\/td\u003e\n\u003ctd\u003eUp to 300W total board power\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eInterface\u003c\/td\u003e\n\u003ctd\u003ePCIe 4.0 x16\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eNetworking\u003c\/td\u003e\n\u003ctd\u003e2x QSFP-DD 400GbE + 2x Warp 100 bridge for multichip mesh scaling\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eForm factor\u003c\/td\u003e\n\u003ctd\u003eActively cooled PCIe card\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eSoftware\u003c\/td\u003e\n\u003ctd\u003eFully open-source stack (TT-Metalium, TT-NN)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003c\/tbody\u003e\n\u003c\/table\u003e\n\u003ch2\u003eGood to know\u003c\/h2\u003e\n\u003cul\u003e\n\u003cli\u003eIt is not a drop-in replacement for a mature CUDA or ROCm stack running mainstream, off-the-shelf inference frameworks.\u003c\/li\u003e\n\u003cli\u003eTenstorrent’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.\u003c\/li\u003e\n\u003cli\u003eExpect to engage with the software stack directly rather than installing a framework with pip and getting instant parity with an NVIDIA card.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003cp\u003eChoose 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.\u003c\/p\u003e","brand":"Tenstorrent","offers":[{"title":"Default Title","offer_id":58232538071422,"sku":"TT-WORMHOLE-N300","price":1399.0,"currency_code":"USD","in_stock":true}],"thumbnail_url":"\/\/cdn.shopify.com\/s\/files\/1\/1022\/4468\/0062\/files\/ai-wormhole-n300-scaled_7db1e3dc-2c7c-4af3-9389-8fc66502ef24.png?v=1792094043"},{"product_id":"tenstorrent-blackhole-p150a-32gb","title":"Tenstorrent Blackhole p150a (32GB)","description":"\u003cp\u003eThe 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. It packs more Tensix cores, more on-chip SRAM and faster memory into a single-ASIC design, backed by a fully open software stack. It suits researchers, kernel engineers and AI developers who want serious on-card compute in a desktop- or workstation-ready package.\u003c\/p\u003e\n\u003ch2\u003eWho it's for\u003c\/h2\u003e\n\u003cul\u003e\n\u003cli\u003eAI research teams building or fine-tuning custom model architectures\u003c\/li\u003e\n\u003cli\u003eKernel engineers who want to hand-optimise compute graphs at a level GPU vendors do not expose\u003c\/li\u003e\n\u003cli\u003eRISC-V-focused developers who want general-purpose cores sitting beside the AI compute fabric\u003c\/li\u003e\n\u003cli\u003eTeams planning a multi-card cluster that scales bandwidth beyond standard PCIe-only multi-GPU setups\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003ch2\u003eKey features\u003c\/h2\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cstrong\u003e120 Tensix cores:\u003c\/strong\u003e Tenstorrent’s core compute unit, each built around a cluster of small RISC-V processors rather than a traditional GPU shader pipeline, for fine-grained control over how work is scheduled and executed.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e16 “big” RISC-V cores:\u003c\/strong\u003e general-purpose processors on the die for control, orchestration and host-side logic, so latency-sensitive or highly custom pipelines make fewer round-trips to the CPU.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e180MB of on-chip SRAM:\u003c\/strong\u003e roughly 1.5MB per Tensix core, a large local pool that keeps data close to the compute units and eases pressure on off-chip DRAM bandwidth.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e32GB GDDR6 on a 256-bit bus:\u003c\/strong\u003e 16GT\/sec for 512GB\/s of memory bandwidth, enough room for meaningfully large local models and datasets.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eUp to 1.35GHz AI clock:\u003c\/strong\u003e delivers 664 TeraFLOPS at BLOCKFP8 precision, a substantial step up from Wormhole.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e4x QSFP-DD 800G ports:\u003c\/strong\u003e passive high-speed networking that links multiple Blackhole cards to pool memory and scale compute across a multi-card system.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003ePCIe 5.0 x16 host interface:\u003c\/strong\u003e the latest-generation host interconnect for maximum throughput between card and system.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eFully open-source software stack:\u003c\/strong\u003e the open compiler and runtime approach (TT-Metalium, TT-NN) that defines Tenstorrent’s product line, with nothing hidden behind proprietary drivers.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003ch2\u003eSpecifications\u003c\/h2\u003e\n\u003ctable\u003e\n\u003cthead\u003e\u003ctr\u003e\n\u003cth\u003eSpec\u003c\/th\u003e\n\u003cth\u003eDetail\u003c\/th\u003e\n\u003c\/tr\u003e\u003c\/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd\u003eTensix cores\u003c\/td\u003e\n\u003ctd\u003e120\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eRISC-V cores\u003c\/td\u003e\n\u003ctd\u003e16 (“big” RISC-V, on-die)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eOn-chip SRAM\u003c\/td\u003e\n\u003ctd\u003e180MB (~1.5MB per Tensix core)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eMemory\u003c\/td\u003e\n\u003ctd\u003e32GB GDDR6, 256-bit bus @ 16GT\/sec\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eMemory bandwidth\u003c\/td\u003e\n\u003ctd\u003e512GB\/s\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eAI clock\u003c\/td\u003e\n\u003ctd\u003eUp to 1.35GHz\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eCompute performance\u003c\/td\u003e\n\u003ctd\u003e664 TeraFLOPS (BLOCKFP8)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003ePower\u003c\/td\u003e\n\u003ctd\u003e300W total board power, 1x 12+4-pin (12V-2×6) connector\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eInterface\u003c\/td\u003e\n\u003ctd\u003ePCIe 5.0 x16\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eNetworking\u003c\/td\u003e\n\u003ctd\u003e4x QSFP-DD 800G (passive) for multi-card scale-out\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eDimensions\u003c\/td\u003e\n\u003ctd\u003e42mm x 270mm x 111mm, actively cooled\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eSoftware\u003c\/td\u003e\n\u003ctd\u003eFully open-source stack (TT-Metalium, TT-NN)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003c\/tbody\u003e\n\u003c\/table\u003e\n\u003ch2\u003eGood to know\u003c\/h2\u003e\n\u003cul\u003e\n\u003cli\u003eThe trade-off is software maturity versus openness. TT-Metalium and TT-NN are open, actively developed and show how your workload maps to silicon.\u003c\/li\u003e\n\u003cli\u003eThe ecosystem of pre-optimised model support is smaller than CUDA’s or even ROCm’s. Expect to work with Tenstorrent’s programming model directly rather than getting instant, zero-effort parity with a mainstream GPU framework.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003cp\u003eChoose the p150a if you value an open stack you can read and modify at every layer and are willing to invest in learning it. If you want mainstream frameworks to work with no effort, a conventional GPU is the easier route.\u003c\/p\u003e","brand":"Tenstorrent","offers":[{"title":"Default Title","offer_id":58232538104190,"sku":"TT-BLACKHOLE-P150A","price":1399.0,"currency_code":"USD","in_stock":true}],"thumbnail_url":"\/\/cdn.shopify.com\/s\/files\/1\/1022\/4468\/0062\/files\/ai-blackhole-p150_589b92da-0b08-4c15-9da9-37c1c6bd6696.png?v=1792094044"}],"url":"https:\/\/computahardware.net\/collections\/tenstorrent.oembed","provider":"ComputaHardware","version":"1.0","type":"link"}