Tenstorrent Blackhole p150
The Blackhole p150 is Tenstorrent's single-chip PCIe 5.0 x16 AI accelerator card: one Blackhole ASIC with 120 enabled Tensix cores, 180 MB of on-chip SRAM, 32 GB of GDDR6 at 512 GB/s and a 300 W board power, sold in an actively cooled p150a for workstations and a passively cooled p150b for server chassis. Its distinguishing feature is four 800G QSFP-DD ports on the card that link Blackhole cards directly to each other, so a multi-card mesh does not ride the host PCIe bus. At $1,399 list it is the cheapest way to buy 32 GB of accelerator memory with a card-to-card fabric, but it runs Tenstorrent's own open-source stack (TT-Metalium, TT-NN), not CUDA, and its 512 GB/s memory bandwidth caps single-stream decode speed well below current NVIDIA cards.
- 512 GB/s of GDDR6 is the decode ceiling. For single-user token generation a used RTX 3090 (936 GB/s) can still be faster, reported; the p150's case is memory per dollar and the on-card mesh, not raw single-stream speed.
- The 664 TFLOPS figure is BLOCKFP8, Tenstorrent's block floating-point format, and is the only precision the vendor publishes. Do not compare it directly with NVIDIA dense FP8 or FP4 numbers without a measured benchmark on your model.
- The four QSFP-DD ports are the real differentiator: cards talk to each other over 800G links without crossing host PCIe, so a four-card box scales tensor-parallel work better than four PCIe-only cards would. They only connect to other Blackhole cards; they are not a general Ethernet NIC.
- Active Tensix cores dropped from 140 to 120 with firmware v19.5.0 and the published peak fell from roughly 774 to 664 TFLOPS; Tenstorrent says typical workloads lose only about 1 to 2%, reported. Check which firmware any benchmark you read was run on.
- This is not a CUDA part. Every model needs a TT-Metalium/TT-NN bring-up, and published Blackhole model coverage trails Wormhole. Treat throughput claims as preliminary until you run your own model on the current release.
What is the difference between the p150a and the p150b?
Cooling only. Both have the same Blackhole ASIC, 32 GB GDDR6 at 512 GB/s, four 800G QSFP-DD ports and 300 W board power. The p150a has an active fan for workstations; the p150b is passive and needs at least 30 CFM of chassis airflow in a server. Both list at $1,399.
How much does the Blackhole p150 cost?
$1,399 list from Tenstorrent's own store for either variant. The p100a, which has 28 GB and no QSFP-DD ports, is $999. QSFP-DD 800G cables are $200 each, with 50% off one cable when you buy two or more p150 cards.
Does the p150 run CUDA or PyTorch models directly?
No CUDA. It runs Tenstorrent's open-source stack (TT-Metalium and the TT-NN library), and each model needs support in that stack. Check Tenstorrent's model bring-up tables for your exact model and software release before buying.
How does the p150 compare with an NVIDIA consumer GPU for LLM inference?
It has more memory than most consumer cards at its price (32 GB) and a card-to-card fabric they lack, but 512 GB/s of memory bandwidth is lower than an RTX 3090 or 5090, so single-stream decode is slower. Its strength is scaling across several cards over the QSFP-DD mesh; independent benchmarks are still scarce.