Google May Make Disruptive Migration To AMD EPYC In Cloud Data Centers
This entry was posted on July 30, 2019.
As that day comes, Lynx Equity Research analysts said they have heard "rumblings" that Google is not happy with Intel's server platform. This prompted Seeking Alpha to posit in a headline, "Google data center switching to AMD?"
It's not clear how much legitimate weight this rumor carries, though Lynx Equity Research believes that Google-specific server boards are being made for AMD's EPYC CPUs, based on its research into the hardware supply chain.
Will it actually happen, though?
"'Switching' is probably the wrong word (in the SA headline), because it's not like they're going to retire thousands of Xeon servers overnight, but adopting EPYC moving forward seems completely plausible, especially with PCIe 4 and nearly double the cores per 2P platform," HotHardware Managing Editor Marco Chiappetta said.
"Totally agree with Marco’s quick assessment here, core count and thread count per socket is a significant advantage for AMD currently. Also, improved IPC throughput and clock scaling of Zen 2 EPYC (Rome) is significantly better than previous gen AMD. However, sever arch migrations have a long gestation period—think 9–12 months and even then, legacy technology will remain. Regardless, AMD getting a nod (if this rumor is true by the analyst) from Google Cloud is huge," HotHardware Editor-in-Chief Dave Altavilla added.
The advantage that Dave references is one that AMD has been keen to note in the lead up to the launch of its next-gen EPYC processors. Back at AMD's Next Horizon event, company CEO Dr. Lisa Su noted that the top-end Rome part will be a 64-core/-128-thread beast that is socket-compatible to the previous gen EPYC platforms, as well as its next-gen Milan server platform (which supports PCIe 4). So, the claim is 2X overall performance per socket and 4X the floating point performance.
AMD also teed up a demo of a dual-socket Intel Platinum 8180M server (56 cores total, 28 per socket, 112 threads) competing against a single-socket AMD Rome EPYC 64-core server C-Ray, which is a ray tracing benchmark designed to showcase floating point CPU performance. As one would expect (given the advantage in cores and threads), the single-socket EPYC part edged out Intel's Platinum-class dual-socket server, as you can see in the video embedded above.
When enabling NAMD optimizations, Intel's own benchmarking data shows its Xeon Platinum 8280 setup scoring 30 percent higher, at 12.65ns. AMD still wins in this scenario, but the difference is far less dramatic, especially when factoring in the core and thread disadvantages that Intel's platform is working with in this comparison.
True comparisons by independent reviewers (such as HotHardware) will have to wait until AMD ships its next-gen EPYC processors. That said, the fact that Google making a migration to AMD can even be a plausible rumor speaks volumes about how far AMD has come.