Americas

  • United States

Intel reshuffles Xeon lineup

News Analysis
Mar 20, 20203 mins
Computers and Peripherals

New Intel Xeon processors with technology for AI training will only come on four- and eight-socket servers. Single and dual-socket versions of Cooper Lake CPUs have been canceled.

xeonw3175x universal right
Credit: Intel

Intel is adjusting plans for its upcoming Cooper Lake generation of Xeon processors by reducing the number of processors it plans to launch and reducing the server types supported.

The company still plans to launch new Xeon processors under the Cooper Lake codename, but it does not plan to support one- and two-socket servers with that product line. It will now focus on high-end four- and eight-socket servers sold under the Cedar Island codename.

Cooper Lake’s big claim to fame is that it would support bfloat16 instructions inside the AVX-512 vector units. The bfloat16 standard was developed by Google for AI training and is extremely technical (here’s a deep dive on the functionality if you wish to learn more).

For end users, the bfloat16 standard is very important to machine learning algorithms, particularly AI training. While virtually all AI training is done with 32-bit floating point calculations, the training doesn’t need 32-bit FP math; 16-bit will do just fine. And 32-bit uses a lot more time and power than a 16-bit calculation.

Therefore, bfloat16 enables AI training with greater speed and power efficiency without sacrificing accuracy. That makes it a big deal.

So instead of the planned release of Cooper Lake processors across two different server platforms – Whitley for one- and two-socket processors, and Cedar Island for four and eight-socket processors – Cooper Lake gets Cedar Island, and the upcoming Ice Lake generation of 10nm processors will go on the Whitley platform for single- and dual-socket platforms.

“Intel constantly evaluates our product roadmaps to ensure we are positioned to deliver the best silicon portfolio for data center platforms,” said an Intel spokesperson in an emailed statement. “Given the continued success of our recent expansion of 2nd Generation Xeon Scalable products and customer demand for our upcoming 10nm Ice Lake processors, we have decided to narrow the delivery of our Cooper Lake products that best meets our market demand.”

Ice Lake is built on a smaller manufacturing process than Cooper Lake – 14nm vs. 10nm – but Cooper Lake has more cores (48 vs. 38 in Ice Lake), and more importantly, Cooper Lake has bfloat16.

So Intel is positioning Ice Lake on the lower end of the Xeon spectrum and Cooper Lake on the high end, which makes sense, says Jim McGregor, principal analyst with Tirias Research.

“They are trying to design for different workloads. There has never been a single Xeon design that has been the same across all apps,” he says. “With deep learning, it’s about density. While most of the learning is done on GPUs or FPGAs, the vast majority of inference is done on CPUs, which is where bfloat16 is used.”

Cooper Lake is expected to be released in the first half of this year, and Ice Lake is due later in the year.