No summer doldrums here. As part of the AMD Epyc Rome launch event in San Francisco today, Lenovo announced two new single-socket servers, the ThinkSystem SR635 and SR655, built with AMD’s new Epyc 7nm 7002 CPU, with up to 64 cores and 128 PCIe lanes.
The new Lenovo servers, set to begin shipping by the end of the month, are billed as having two-socket performance in a 2U, single-socket design, delivering twice the throughput and four times the floating point capability of ThinkSystems equipped with previous-generation processors. The systems also have up to 9x PCIe Gen4 slots and memory speeds up to 3200MHz, Lenovo said. The servers support up to six single-wide GPUs (or up to three double-wide GPUs), and the SR655 can support up to 32 NVMe solid-state drives.
At a pre-announcement roundtable discussion last night with executives from Lenovo and AMD, Scott Aylor, AMD Datacenter Solutions Group’s corporate VP/GM, said the new servers are “able to attack the entire market that is two-socket today. We believe that with Epyc Rome (7002) there’s probably not a two-socket application that we aren’t within a stone’s throw of, regardless of the Intel CPU you buy.”
A bold statement, no doubt, but Charles King, president/principal analyst at industry watcher Pund-IT, told us today that, “I think Aylor’s statement is fair. While these initial offerings are single socket servers, it’ll be interesting to see how second gen Epyc matches up against comparable Intel Xeon silicon in more complex systems. From what I can see, the new Lenovo servers should match up well against the company’s comparable Xeon-based offerings. Customers who have long seen AMD as little more than a low cost alternative to Intel will need to reconsider that assessment.”
Said Patrick Moorhead, president/principal analyst, Moor Insights & Strategy, “I am impressed with Lenovo’s time to market with an optimized platform for the second gen Epyc, and even more surprised to see some of the world record performance numbers the company racked up. Lenovo appears serious about driving some real volume with AMD, something I have never seen on its server line.”
Moorhead said early indications from AMD and its OEM partners indicate that second-gen Epyc is performing “exceptionally well and has advantages in many different workloads, but not all workloads. AMD looks strong in Hadoop RT analytics (AMD says world record), Java throughput (AMD says 83 percent better), fluid dynamics (AMD says 2X better), and virtualization (AMD says up to 50 percent lower TCO). Intel will likely have advantages on low latency ML inference workloads that take advantage of Intel’s DLBoost instructions. Intel will also look very good in in-memory database workloads utilizing Optane DC.”
At last night’s event, AMD – naturally enough – focused on processor performance while Lenovo highlighted the servers’ overall capabilities. Doug Fisher, COO for Lenovo’s DCG and SVP, DCG Solutions and Segments, said the server market has evolved from a focus on component technologies to the hardware’s ability to support workloads and use cases.
“Customers in my view…do not sit there and say I need to buy a Lenovo system with AMD in it,” Fisher said. “They start with what’s the business problem…, (they) usually start at some workload, some software capability and then it generates into some piece of hardware. What Lenovo does is we spent a lot of money in differentiating that hardware – which is reliability, performance. AMD does the same thing. AMD is adding capabilities and features that are unique, whether it’s PCIe Gen4, whether it’s their security features.”
Workloads suited to the new servers include video security, software-defined storage and network intelligence, as well as support for virtualized and edge environments. The company cited the emerging IoT market that places a premium on use cases where more computing power is needed in confined spaces, using less energy, such as video and image analysis for public safety in smart city, campus and mass transit environments.
Kamran Amini, executive director of Lenovo Enterprise Business Group added that he doesn’t “believe an x86 server is an x86 server – there is engineering that creates differentiation.”
The companies also emphasized that they view the servers as gateways to a range of market segments, including high-performance computing-class use cases, where single-socket servers can deliver licensing advantages.“In our view,” Amini said, “with AMD, they are delivering massive amounts of cores per socket, memory speed is going to be key, especially for memory-intensive applications. As for I/O — I’m getting a bunch of basically free lanes, I don’t have to add a PCI adaptor card to create more lanes, which gives me better economics as well. But then on top of that, we’re going to have the most dense NVMe server in the market with the (SR)635 having 16 NVMe drives – so density meaning capacity as well – leveraging the two-and-half-inch drives. 32 (NVMe drives) for the (SR)655. The view was how do we take where we know AMD is going to create differentiation and value and then be able to extend that beyond based on the application of workloads running on.”
“As it relates to HPC’s affinity for single socket, there are a number of workloads that tend to be more core licensing-based than socket-based licensing, they tend to be lower core count, high frequency-oriented workloads,” said AMD’s Aylor, citing, for example, electronic design automation (EDA) applications.
“So think about things…where you’ve got a Cadence license, a Synopsys license; and those quite frankly, if you look at the licensing cost per core, it’s probably the best thing in the world to get the highest performance CPU you can in that socket because you’re going to pay a lot more in licensing than you are for the hardware. So we see a lot of opportunity in things like financial services, and EDA for single socket because the performance that is needed and the heavy licensing costs that are there from a core technology perspective are very relevant. So it’s very different from the scale-out, memory bandwidth-oriented things that you would expect to see in two-socket, but there are some important sub-segments where one socket matters.”