Intel acquired Habana Labs, an Israel-based developer of programmable deep-learning accelerators, in December 2019 for approximately $2 billion. Habana’s Gaudi line became part of Intel’s data-center AI portfolio. Gaudi accelerators handle machine-learning training and inference, while their chips are manufactured by a foundry rather than at Intel’s Kiryat Gat plant.
What Gaudi accelerates
Training adjusts a model’s parameters using data; inference runs a trained model to produce results. Both can involve large numbers of parallel calculations and frequent exchanges of data among accelerators. Gaudi combines compute engines, high-bandwidth memory and high-speed networking for these data-center tasks. Intel’s 2019 acquisition announcement describes Habana as a developer of programmable deep-learning accelerators for data centers and says the deal strengthened Intel’s AI portfolio.
Intel’s July 2025 Gaudi 3 technical paper specifies 1.8 petaflops of FP8 and BF16 compute, 128 GB of HBM2e memory and 3.7 TB/s of memory bandwidth. Those figures are tied to named numerical formats and product specifications; they are not a promise of identical speed for every model or workload. The paper presents Gaudi 3 for generative AI, large language models, diffusion models and other AI workloads.
Design and manufacturing are different jobs
Intel’s paper says Gaudi 3 uses TSMC’s 5-nanometer process. Intel designs the accelerator and its system, while Taiwan Semiconductor Manufacturing Company fabricates the silicon. This distinction is central to the global chip supply chain: an Israeli-developed product can depend on manufacturing capacity located abroad.
Habana began as an Israeli company, and its acquisition gave Intel a specialized accelerator architecture and engineering base. Intel’s acquisition notice established the Israeli origin and product focus; its later product paper supplies the Gaudi 3 specifications. Read about Intel’s Haifa and Kiryat Gat operations to see how processor development and wafer fabrication also take place at separate sites.