Advanced Computing in the Age of AI | Friday, March 29, 2024

Google Embraces AMD Epyc Rome CPU for Cloud, Internal Workloads 

Google’s Bart Sano joins AMD’s Lisa Su on stage at the Rome launch (Aug. 7, 2019)

This week’s big tech news – AMD’s release of the Epyc Rome CPU, the industry’s first 7nm server chip – got a major boost when Google confirmed that it now uses the new processor for internal workloads and will soon make it available to Google Cloud customers. In AMD’s struggle to take market share from Intel in the lucrative data center server processor market, it doesn’t get much better than a FAANG hyperscaler win.

Rumors of Google’s AMD commitment surfaced last week from Lynx Equity Limited analysts, who heard "rumblings" from within Google’s supply chain providers that the company was not happy with Intel’s server platform. The decision stems from Epyc’s 7nm price/performance advantage over Intel Xeon Scalable 14nm chips, with 10nm server CPUs from Intel still more than a year away. Further, Epyc is priced at under $7,000 and has over 2x the cores of Xeon Platinum 8280, priced at roughly $10,000.

Google said the 2nd Gen Epyc processor-powered virtual machines it offers its public cloud customers “will be the largest general-purpose VMs we’ve ever offered.”

“We believe that many general-purpose workloads, including back-office applications and web servers, will see price-performance improvements on the new AMD VMs compared to their current configurations,” Google said in a blog. “Big compute workloads driven by memory bandwidth such as financial simulations, reservoir analyses, and weather modeling, can take advantage of full-socket VM sizes that provide up to 60 percent better platform memory bandwidth than existing instances.”

Google said the AMD-powered VMs will be available later this year.

AMD announced the 2nd Gen Epyc chip at a gala event in San Francisco yesterday. The new AMD processors feature up to 64 “Zen 2” cores per SOC, deliver up to 23 percent more instructions per clock (IPC) per core on server workloads and up to 4 times more L3 cache compared to the previous generation, said AMD. The company also reported that it holds 80 world records, 15 of them in HPC, while delivering an estimated 25-50 percent lower TCO than competitive offerings.

The Google decision may bode well for AMD in the data center processor market. Industry observer Patrick Moorhead, Moor Insights & Strategy president and principal analyst, pointed to the risk averseness of major IT buyers, who typically don’t commit to early product versions.

“AMD gained low, single-digit share with 1st Gen Epyc but I expect the company to gain more share with 2nd Gen EPYC with CSPs, enterprises and HPC,” he said. “Enterprises don't mass deploy any first gen product, they didn't deploy 1st Gen Epyc, but they will deploy the 2nd Gen Epyc.”

FAANG companies like Google purchase IT gear at such a scale that each of them, to a degree, is a market unto themselves. This trend will grow in significance as enterprise workloads increasingly move to public cloud services providers, which run enormous data centers all over the world.

Google said its AMD VMs are available with 2.25 Ghz base frequency, 2.7Ghz all-core-turbo frequency, and 3.3Ghz single-core turbo frequency utilizing Epyc scalability, which starts at 2 vCPUs and scales up to over 200 vCPUs. The VMs will support RAM-to-vCPU ratios from 1 to 8.

EnterpriseAI