Intel Kicks Off Vision 2022 With Habana Gaudi 2, Greco, 12th Gen Core HX, IPUs

Intel is kicking off their inaugural "Intel Vision" event today in Texas by making several prominent hardware and software announcements.

Intel is using Vision 2022 to formally launch their 12th Gen "Alder Lake" Core HX processors for mobile workstations. The 12th Gen Intel Core HX processors feature up to 16 cores and up to a 5GHz clock speed with the flagship Core i9 12900HX.

Intel says they are now shipping their initial SKUs for 4th Gen Intel Xeon Scalable "Sapphire Rapids" processors. Shipping are "the first of many SKUs" while they will be ramping up through the remainder of 2022. Details on these initial SKUs though are still light and we'll learn more later in the year and hopefully when we finally have our hands on Sapphire Rapids for benchmarking.

We've already been talking about Intel's Arctic Sound M (ATS-M) server GPU for some time, mostly in the context of the open-source Linux driver support. Intel is using Vision 2022 to talk up Arctic Sound M's AV1 hardware encode capabilities, a performance target of 150 TOPS, and the open-source oneAPI software stack support. The formal ATS-M launch now is pegged for Q3.

On the data center side Intel also is laying out more of their Infrastructure Processing Unit (IPU) details at Vision 2022. This year has their "Oak Springs Canyon" FPGA platform and "Mount Evans" ASIC while 2023~2024 will see the Hot Springs Canyon platform and Mount Morgan ASIC with 400GB/s connectivity, up from 200GB/s. For 2025/2026 they have an unannounced, next-gen offering that will offer 800GB/s capabilities.

Intel continues talking up their open-source software ecosystem for IPUs and all of their other products for that matter. Great seeing their open-source software efforts continue full-speed under Pat Gelsinger's leadership.

The Habana Gaudi2 is being announced as their next-generation AI processor for training and inference. Intel acquired Habana Labs back in 2019 and the Gaudi2 is now ready as their next-gen AI accelerator. Intel is promoting the Gaudi2 as delivering two times better AI training performance than current NVIDIA A100-based offerings. Gaudi2 moves down to 7nm manufacturing, goes from 8 to 24 TPCs for media decode and processing, sports 96GB of HBM2e memory and a 48MB SRAM cache, sports 24 x 100 GbE networking, and has a 600 Watt TDP.

The Habana Greco has also been announced as their new offering optimized for deep learning inference efficiency. Greco will be shipping to customers in the second half of the year. Going from the existing Habana Goya to the Greco means 16nm down to 7nm manufacturing, 40 GB/s DDR4 memory up to 204 GB/s LPDDR5, various new hardware capabilities, and being a single-slot HHHL card with a 75 Watt TDP compared to a 200 Watt draw with Goya.

Intel Habana Gaudi2 and Greco hardware will continue using the SynapseAI software suite. This goes along with their open-source Habana Labs AI kernel driver and their now open-source user-space software components, including the compiler. Originally the Habana Labs user-space components were closed-source and that is what caused much friction with the upstream Linux kernel developers given their kernel driver has been open-source and upstream. Last year Intel opened up the SynapseAI Core code to address those prior binary blob issues.

Over on the software side, Intel is announcing "Project Apollo" as a collection of more than 30 open-source AI solutions designed for making AI more accessible for on-prem, cloud, and edge environments.

Overall some exciting announcements and will be eager to get my hands on some of the hardware to put it in the lab for Linux support analysis and benchmarking.