AMD’s New Epyc Server Brand to Take on Deep Learning

Discussion created by kingfish on May 18, 2017
Latest reply on May 18, 2017 by black_zion

AMD's New Epyc Server Brand to Take on Deep Learning - ExtremeTech

"We’ve known the codename for AMD’s Ryzen server CPUs, Naples, but the company unveiled the actual brand name today. Henceforth, AMD’s server CPUs will be sold under the Epyc brand. It’s pronounced like the ordinary word, and as brand names go…maybe it’ll grow on people? That’s pretty much the best we’ve got right now. Opteron was a stronger product name in our personal opinion, but ultimately, customers aren’t going to care what it’s called, they’ll care how it performs.

One point AMD made from the beginning of the event is that it focused on data centers when it brought Zen to market. We haven’t seen much in the way of server benchmarks on the core yet, but Ryzen 5 and 7 have been at their strongest when competing against Intel in workstation workloads like video editing and 3D rendering. Server workloads tend to be their own unique animal, but AMD enjoyed tremendous success with Opteron when it aimed Athlon 64 at the data center first, consumer workloads second. It’s not surprising to see company doubling down on this strategy again."

The company’s Epyc processors will offer up to 32 cores with 128 lanes of PCI Express 3.0 connectivity and eight memory channels. This combination of massive memory bandwidth and core counts could give AMD a leg up on Intel in three key areas. First, AMD can offer an enormous amount of memory bandwidth on a single-socket system with all eight of its memory channels. Intel, in contrast, is limited to just four channels. Granted, Intel’s memory controller efficiency might prove to be higher than AMD’s, but eight channels is hard to argue with, particularly in a one-socket system, and the two-socket comparison is pretty strongly tilted for AMD.

Second, AMD is purposefully eschewing Intel’s habit of locking off certain capabilities to specific SKUs. Intel’s Xeon lines are subdivided by capability — some SKUs are for single-socket systems, some support two sockets, some range up to four sockets. AMD is simply declaring that all versions of all Epyc chips will have the same I/O, the same memory channels, and the same capabilities.

Third, AMD wants to position Epyc as an ideal match for machine learning and artificial intelligence workloads. To that end, it’s built a single-socket CPU with enough I/O connectivity to handle multiple Radeon Instinct cards. AMD noted that in other servers, customers are required to buy multi-socket systems simply to have enough connectivity to run a full complement of GPUs in them, as shown above. The 128 lanes of PCI Express connectivity that Epyc offers, in contrast, can drive six cards (16 lanes each) from one socket, with 32 lanes of PCI Express 3.0 left over for intra-chip communication.

One point of clarification here: Unlike Intel, which offers 40 PCI Express 3.0 lanes in a single socket and 80 lanes in a dual socket, AMD offers 128 lanes in both a 1S and 2S configuration. In the two-socket system, 64 lanes of PCIe 3.0 are dedicated to intra-chip communication, with each CPU offering 64 lanes of chipset I/O.