Back to Terafab AITFAB-11
Terafab AI · hardware
TF-1 Accelerator
Custom AI training and inference silicon
Process node
3nm
Memory bandwidth
3.2 TB/s
Peak compute (FP8)
2 PFLOPs
Price
From $18,000 / unit - enterprise quote
Overview
TF-1 is Terafab's first-generation AI accelerator, designed in-house and fabricated on a 3nm process specifically to run Grok training and inference workloads more efficiently than off-the-shelf GPUs. It's deployed inside Colossus alongside third-party accelerators, with the group's roadmap shifting more capacity to TF silicon each generation.
What sets it apart
Purpose-built for Grok
TF-1's architecture is tuned specifically for the model shapes xAI trains, trading general-purpose flexibility for efficiency on known workloads.
Lower cost per token
Vertical integration from design to fabrication cuts the cost of training and inference compute compared to renting third-party GPU capacity.
Full specifications
Specifications
Process node
3nm
Memory bandwidth
3.2 TB/s
Peak compute (FP8)
2 PFLOPs
Interconnect
Custom high-bandwidth fabric
TFAB-11 · Terafab AI