Something EPYC is missing from mainstream servers

HPE Cloudline CL3150 server

The major server vendors have all now officially announced a refresh of their portfolios, most timing this to coincide with the official launch of Intel’s Xeon Scalable processors this week.

HPE detailed its Gen 10 wave of servers last month, but Dell EMC, Lenovo and Cisco all saved much of the detail until Intel’s official announcement regarding its new server processors, based on the Skylake architecture.

This is no real surprise, as we’ve got used to the vendors tying their refresh cycles to the availability of new processor platforms.

But wait, where are the systems based on AMD’s EPYC processor platform, a chip designed to take on Intel’s Xeon processors? At the EPYC launch in Austin, Texas last month, AMD disclosed a number of vendors supporting their new chips, which included Dell EMC, HPE and Lenovo.

I asked Lenovo where its AMD-based systems are, and the firm was somewhat less than forthcoming. One product manager stated that he had no knowledge of any EPYC models, while the official line from the firm’s press office is that their servers are Intel-based.

So, I checked back to AMD’s announcement of the EPYC server chips, and sure enough, there is an endorsement from Lenovo:

“The AMD EPYC processors present unique opportunities for our customers to lower Total Cost of Ownership via an unprecedented balance of cores, memory bandwidth, and I/O. We are excited to collaborate with AMD and several global Hyperscale customers to develop and deploy single socket and dual socket EPYC-based servers,” said Paul Ju, vice president and general manager, Lenovo Global Hyperscale Business.

It would seem that if Lenovo is using AMD EPYC chips in systems, these are being supplied by a different business unit that targets the hyperscale market, while its enterprise servers are entirely Intel-based.

Lenovo is not the only vendor taking this path; HPE’s sole EPYC server disclosed so far is the Cloudline CL3150 (pictured), an ultra-dense rack-mount system targeting software-designed storage clusters, again for the hyperscale and service provider markets.

Many of the other brands lined up by AMD are well-known white-box vendors such as Supermicro, Tyan, Inventec and Wistron, which also sell into the hyperscale market.

Meanwhile, Dell EMC gave a supporting presentation at AMD’s event in Austin, yet its 14G announcement makes no mention of EPYC-based systems, nor are any listed on its site.

The situation thus seems to be that the Tier 1 vendors are all steering away from offering EPYC-based systems to compete directly with Intel’s Xeon platform, except in the hyperscale arena, where factors such as up-front cost and performance per watt override other considerations.

Is this because the big vendors not want to risk the wrath of Intel by putting up Xeon and EPYC systems against each other, or is it that their analysis of AMD’s chips has convinced all of them that it is ideally suited to a different role than that of Intel’s?

The reality is that it could be either reason, or both. Or to put it another way, vendors such as HPE enthusiastically backed AMD’s earlier Opteron server chips, only to see the company fail to keep up with Intel’s rapid pace of development, and so the sizable fraction of the server market that AMD managed to grab for itself slowly declined and customers drifted back to Intel-based systems as their servers reached end of life and were replaced.

Lenovo, HPE and Dell EMC are likely taking the approach that they won’t rush to offer AMD-based enterprise servers immediately, and are waiting to see if there is customer demand for them. With server shipments to enterprises reportedly declining because of the growth in cloud services, this may never materialise.

Then again, perhaps AMD was wise to develop a chip that is ideally suited to high-density servers for the service provider market: this is the area that is showing the strongest growth at the moment, after all. With up to 32 cores, 8 memory channels and 128 PCIe lanes, AMD can claim with some justification that a single-socket EPYC server can do the duty of a traditional two-socket system, at much lower cost.

UPDATE: I asked Dell for details of its EPYC-based servers, and this is the response I received:

Yesterday Dell EMC launched the first wave of its 14th Generation of PowerEdge servers aligned to Intel processors, PowerEdge servers based on the AMD EPYC processors will be available in the second half of 2017.