Intel announced new chips optimized for AI workloads and datacenter acceleration.

At its inaugural Intel Vision event, Intel Corp. expanded its data center and AI product portfolio with several new processors and accelerators. The company unveiled advanced chips from Habana Labs, its AI chip subsidiary acquired for $2 billion in 2019, alongside new infrastructure processing units (IPUs) and graphics cards designed for enterprise computing workloads. Habana Labs, which was acquired to develop AI chips for data centers with capabilities in both model training and inference, announced next-generation versions of its training and inference processors.

Intel introduced Gaudi 2, the latest training chip from Habana Labs, which represents a significant upgrade from previous-generation silicon. Moving from a 16-nanometer to a seven-nanometer manufacturing process, Gaudi 2 includes 24 Tensor Cores—tripling the eight Tensor Cores in the predecessor—and has doubled its SRAM memory and tripled its HBM2E memory. According to Intel's internal benchmarks, Gaudi 2 delivers twice as much throughput as Nvidia Corp.'s flagship A100 graphics processing unit for training popular machine learning models, and offers up to three times as much throughput as Habana Labs' first-generation training chip. Habana Labs Chief Operating Officer Eitan Medina stated: "Compared with the A100 GPU, implemented in the same process node and roughly the same die size, Gaudi 2 delivers clear leadership training performance as demonstrated with apples-to-apples comparison on key workloads. This deep-learning acceleration architecture is fundamentally more efficient and backed with a strong roadmap."

Alongside Gaudi 2, Intel announced Greco, Habana Labs' second-generation inference chip for running trained AI models in production, also upgraded to the seven-nanometer process. Greco features an onboard media encoding engine for HEVC, H.264, JPEG and P-JPEG formats and has a thermal design power envelope of 75 watts—less than half the 200-watt envelope of its predecessor. Intel also announced plans to expand its IPU portfolio through 2026, starting with Mount Evans, an application-specific integrated circuit co-developed with Google LLC's cloud business that will launch this year and process 200 gigabits of data per second. Intel plans to follow Mount Evans with a more advanced IPU in 2023 or 2024 capable of processing 400 gigabits per second, and another chip by 2026 processing 800 gigabits per second. Sandra Rivera, Intel's executive vice president and general manager of the Datacenter and AI Group, noted: "These IPUs were co-developed with large-scale customers like Google and Microsoft to enable them to customize their infrastructure."

Completing Intel's data center announcement was the Arctic Sound-M graphics card, designed for video content processing and cloud-based game streaming, capable of performing 150 trillion computing operations per second and launching next quarter in two form factors across more than 15 data center systems from Intel hardware partners. Intel also debuted the Intel Core HX series, a line of seven high-end laptop CPUs optimized for professional applications such as computer-aided design, animation and visual effects, with the fastest processor featuring 16 cores split between performance-optimized and efficient designs.