Intel’s Ice Lake CPU for the data center is better late than never, and promises performance advantages across AI, 5G and crypto workloads.
Intel has finally launched Ice Lake, the first CPU family to be built on the integrated device manufacturer’s (IDM) delayed 10nm process technology. Ice Lake, which is the second family of third-generation Xeon Scalable parts, promises a 46% performance increase on common data center workloads, compared to second-generation parts, Intel said.
Ice Lake was announced at CES in January 2019. At that time, availability was scheduled for 2020. Production delays due to a difficult transition from 14nm to 10nm account for this delay, though Intel will be hoping this first device off the new production line puts any doubts about the new 10nm process to bed.
Intel is rapidly ramping up production of Ice Lake. At least 200,000 parts have already shipped for revenue in Q1 2021, including to numerous cloud service providers, more than 50 ODMs and OEMs, at least 15 telecoms equipment manufacturers and 20+ high performance computing (HPC) customers.
Ice Lake CPUs are the second family in Intel’s third-generation Xeon Scalable CPU offering, and the first to use the new Sunny Cove core microarchitecture. The first family of Intel’s third-generation Xeons, Cooper Lake, launched last summer, and is based on the previous 14nm process node. The upcoming fourth generation is named Sapphire Rapids.
Ice Lake is intended to keep up with changing data center workloads which increasingly include AI, crypto and 5G. AI performance for Ice Lake beat second-generation Xeons by 1.74X (batch inference for BERT natural language processing), with slightly lower uplift on image recognition (Mobilenet-V1) and image classification (ResNet50-v1.5). Intel also released internal benchmark scores placing Ice Lake at up to 1.5X the performance of AMD Epyc 7763 CPU (“Milan”) and 1.3X the performance of Nvidia’s A100 GPU (but like all internal benchmark figures, these should be taken with a pinch of salt).
“Since most data scientists don’t run a single AI benchmark, we looked at a recent Kaggle survey, and chose a broad range of twenty popular machine and deep learning models, including both training and inference,” said Wei Li, VP and GM, Machine Learning Performance at Intel. “Not only did we outperform both AMD Milan and Nvidia A100 on the majority of workloads, we also outperformed them across the geomean of all workloads. Even when we don’t outperform a GPU on some deep learning benchmarks […] customers are still choosing Xeon because of the performance and [total cost of ownership] benefits.”
Intel is facing increasing competition from the likes of AMD and Nvidia in the data center. AMD’s Milan processors launched last month. Built on TSMC 7nm, Milan processors offer 10% more instructions per clock cycle vs their predecessors (Rome). Meanwhile, Nvidia’s flagship data center and HPC GPU, the A100, is the one to beat when it comes to dedicated data center AI acceleration.
Countering this competition, Intel stressed that its 3rd gen Xeons are the only data center CPUs with built-in AI acceleration (this is in the form of DL Boost which includes specialized AVX-512 instructions and support for BF16 data, a popular format in AI training. These features were introduced in second-gen Xeons). AI-specific features like these are missing from AMD’s Milan, for example.
Intel also frequently plays its “system” trump card — its portfolio includes dedicated AI accelerators (via Habana Labs), FPGAs (via Altera), Optane memory, Ethernet NICs, etc. These technologies can be bundled together to make a bigger offering for data centers than just CPUs, allowing Intel to offer reduced total cost of ownership.
However, since Cooper Lake was launched last summer, AMD has announced it intends to acquire FPGA maker Xilinx, whose products also include smart NICs. Nvidia, fresh off its purchase of smart NIC supplier Mellanox in 2019, is also in the process of acquiring CPU IP giant Arm. Arm-based data-center CPUs are making headway in HPC (Fugaku supercomputer) and data center (Ampere). And during Arm’s recent v9 architecture launch, the company said it will increase hardware support for AI across its entire portfolio, including CPUs for HPC and data center. Intel’s “system” trump card, therefore, may not be as effective in the coming years as it is today.
This article was originally published on EE Times.