Kamran Amini was an executive at IBM in the mid-2000s when the company first put AMD’s then-relatively new Opteron processors into some of its System x servers. Big Blue was the first big OEM to do so, ahead of Sun Microsystems and well ahead of Hewlett Packard and Dell.
Just a few years earlier, in 2003, AMD jumped into the datacenter with the first Opteron chip armed with the “Sledgehammer” core and began to eat into rival Intel’s dominant position in the server chip space. Opteron was the first X86 server processor that supported 64-bit memory and processing, an enormous step for an IT industry that had no other alternative. Intel’s Xeon processors supported only 32-bit memory and compute because the company was still pushing its (ultimately doomed) Itanium architecture. Organizations were anxious to be able to run 64-bit applications and didn’t relish the idea of having to adapt to a new architecture. Opteron allowed them to stick with the familiar x86 architecture, and AMD soon was able to grab almost 20 percent of the X86 CPU market in the datacenter.
Amini, now vice president and general manager of server, storage, and software-defined infrastructure for Lenovo’s Data Center Group, tells The Next Platform that “Opteron was an excellent product. They changed the dynamic and the architecture and a lot of people moved from Intel to Opteron.” The problem was that AMD couldn’t capitalize on its success. Subsequent generations of Opteron were found wanting, particularly the ill-fated quad-core “Barcelona” release in 2007, which he calls “a complete disaster in execution.” It allowed a well-funded Intel to catch up and pass AMD with new generations of its Xeon chips and AMD’s place in the datacenter faded.
It also made enterprises wary of AMD and its ability to execute on its roadmap, a killer for any vendor in any industry. There may have been a thirst for a competitive product to Intel, but businesses were not going to risk jumping back on the AMD bandwagon.
Epyc Comes To Market
Fast forward to 2017, and AMD releases “Naples,” the first product in AMD’s new family of Epyc server chips based on the vendor’s new clean-sheet Zen microarchitecture and backed by promises from CEO Lisa Su and other executives about high performance, power and cost efficiencies, and a roadmap that was as set in stone as any industry can be. Last year, AMD rolled out the second generation of Epyc, “Rome,” which included, among other features, 64 Zen 2 cores – twice that found in Naples – four times the L3 cache, 24 slots for NVM-Express storage devices, more security features and an enhanced Infinity Architecture.
The capabilities AMD put into Epyc 7002 series Rome processor and the fact that it was on time made a difference with organizations, Amini says. That becomes important not only to challenge Intel, but because other companies – including Arm and its various partners, such as Ampere and Amazon with its Graviton processors – also are making a play for a share of the datacenter chip market.
“Many customers that deployed AMD many years ago with Opteron and saw the failure with Barcelona were a little hesitant,” he says. “The classical enterprise customers were hesitant to jump on the bandwagon of AMD with Naples. What they wanted to make sure of is [that AMD] can deliver on the value proposition of a product’s performance, time to market, and that they could deliver the roadmap. Not just one CPU, but multiple generations. If you look at what AMD’s done, they’ve promised things and they over-delivered on those promises. You saw that happen with Naples. You saw that happen with Rome. They’re extending value prop.”
That allowed AMD to build confidence from customers outside of those that normally lead with new processor technologies, such as HPC organizations and hyperscalers like Google, Amazon and Facebook. They not only jump in because of the leverage they can get in negotiating pricing with vendors, but also because they’re focused on how much pure horsepower they can get, Amini says. Enterprises are different.
“The commercial enterprise guys are very price-performance sensitive,” he says. “They have to be bought into the capability that you’re going to bring and the longevity, because every time they shift from a major architecture to another architecture, they don’t want to be bouncing around constantly. It costs to them [them to do that], with internal operations, internal validation and everything else.”
Naples was a good start, but Rome really gave enterprises and even OEMs confidence in AMD’s new chips. Businesses evaluated Naples, but Amini says “they wanted to see is their reality to Rome. Rome came out and I think we’re seeing more people being interested in Rome in the enterprise space. As Rome builds confidence, Naples will continue to drive some confidence unless a major disaster happens. The value proposition of AMD is resonating in parts of the market that is looking for pure performance or looking for that one-socket, sweet price-performance level. Of course, Intel’s not going to sit still. It’s a competitive landscape here. The interesting thing is what happens as we look forward.”
Lenovo Goes Two Socket With Rome
Lenovo jumped onto the Rome bandwagon in last summer, launching the one-socket ThinkSystem SR635 and SR655 servers aimed at a broad array of workloads, from analytics and HPC to edge computing and software-defined storage (SDS). Less than a year later, Lenovo this week unveiled two dual-socket servers, the SR645 and SR665, which double up on the core count and target such industries as retail, financial services and manufacturing that can use more capabilities around such tasks as data analytics and transaction processing. They also address the need for improved cost efficiencies. For one, the higher core counts mean that customers need fewer systems to run their workloads and less rack space.
The 2U servers also double the I/O bandwidth, support up to eight Nvidia T4 GPU accelerators and offer up to 40 2.5-inch drives or up to 32 NVM-Express drives, both of which play into dense SDS environments. They also are available via Lenovo’s TruScale program, enabling enterprises to rent the servers, put them into their datacenters and pay for what they use.
Like most of what’s happening in IT today, what it comes down to is data – the amount of it being generated and the varied nature of it.
“In the market we’re seeing really three areas that are driving change for us. We know about data’s massive growth. If you look at the current situation we’re on with coronavirus virus and how you’re trying to determine a vaccine, it’s all collecting a massive amount of data and what you’re doing with the data,” Amini says. “There is the speed of data that’s driving a change in how you think about your IT. Then there’s the variety of data that’s being generated. The variety of data requires insight, requires accelerators to drive more analytics and machine learning and [artificial intelligence]. But data velocity is asking for us to drive massive amounts of throughput capability out of the platform. This is where things like NVM-Express and being able to design a platform that has full maximum memory and speed capability and bandwidth to be able to do in-memory analytics [come into play]. And then the data volume is just essentially, ‘How do I handle that massive amount of capable storage?’ This is about the ability to put 40 SSDs in a single box and a mass of terabytes of footprint inside to handle the amount of data coming in.”
Lenovo’s one-two punch with single-socket and dual-socket servers leveraging the Rome chip lets the OEM cover a wide range of the X86 server market, he says. The value proposition of the one-socket servers was that enterprises and cloud providers could use them to deliver better economics than a dual-socket Intel system in both ownership and licensing costs while providing enough compute, memory and I/O to run the workloads.
“We saw that resonate,” he says. “We’ve seen good success and FSI. We’ve seen some success with even HPC customers. We’ve seen some good success around our channel business because, again, we’re taking different routes to market. Our channel businesses like that one-socket platform. … The two-socket was purely around where we could differentiate in performance. It really was around fine tuning, delivering a balanced platform from a capability and performance standpoint. We’re also expecting where customers running SAP apps – and if you look at AMD in the past, AMD’s not been successful in SAP apps – we can bring further performance enhancements and capability with our new platform to be able to penetrate that space with an AMD architecture beyond what we’ve done in the past.”