Advanced Computing in the Age of AI | Friday, March 29, 2024

Intel Aims Xeon E7 v2 At Big Memory Workloads 

Big companies support thousands of users on their applications, they have big databases behind them, and they therefore need big iron. For those who want to migrate off of RISC and proprietary systems to X86 machines of similar heft, Intel is finally bringing an "Ivy Bridge" update to its high-end Xeon E7 processors to market. It has been a long time since these top-end platforms were refreshed.

One such customer that has an appetite for relatively large Xeon machines is Moody's Investor Service, which has migrated about half of the 2,000 RISC/Unix systems that support its credit rating, research, and risk analysis applications over to a mix of two-socket Xeon E5 and four-socket and larger Xeon E7 machines running Linux. Sajid Khan, global enterprise segment marketing manager in Intel's Datacenter and Connected Systems Group, explained at a pre-briefing ahead of the "Ivy Bridge-EX" Xeon E7 v2 launch that Moody's was able to shave about 10 percent off its operating expenses for systems on the replaced machines in the first year and about 25 percent per year thereafter. Moody's has been doing this migration gradually and carefully over the past two years and presumably with the latest Ivy Bridge Xeon E5 and E7 boxes, the company can finish up the job.

Over the past decade, this same story has played out again and again, which is why the RISC/Unix and proprietary system markets have been declining at such a rapid pace.

Back in 2003, said Khan, X86-based systems accounted for 15 percent of the revenue share of machines with four or more sockets. (This is data from IDC.) Fast forward to the end of 2013, and the X86 platform – which technically includes Xeons as well as Itaniums from Intel and Opterons from AMD – accounted for 46 percent of sales. Starting with the "Nehalem-EX" Xeon 7500 back in 2010, Intel moved into a place where it was not just offering better bang for the buck at the high-end of the server market compared to alternatives, but for the first time delivered equivalent performance to Sparc, Power, and Itanium machines. The "Westmere-EX" chips from 2011 pushed the performance envelope even further, helping Intel and its system partners to take even more share away from the RISC/Unix and proprietary systems makers.

The upshot is that now X86-based processors now account for around 80 percent of shipments of machines with four or more sockets, up from around 50 percent or so a decade ago. The RISC/Unix vendors have had to live on lower volumes, lower prices, and lower margins and are being driven out of this lucrative market as they have largely been pushed out of the two-socket space.

This is ever the cycle. Proprietary minis ate into the mainframe market, and Unix machines ate into both. The X86 market continues to take bites out of all three, and if a sustainable advantage can be found for 64-bit ARM servers, the X86 market may yet get its turn being bitten. So far, this has turned out to be more difficult than many ARM enthusiasts had expected. But remember this: People thought the same thing about Pentium servers in the early 1990s.

intel-xeon-e7-v2-die-shot

There is plenty of pent-up demand for Intel's new Xeon E7 v2 processors, and not just because of ongoing migrations from RISC/Unix and proprietary platforms or the sudden popularity of SAP HANA in-memory databases to run ERP applications faster. (SAP only allows for HANA to run on the top-end Xeon E7 platforms because they have the highest reliability and the highest processor and memory scalability in the X86 space.) Another big reason that there is pent-up demand for Xeon E7 v2 chips is that Intel did not ever deliver a "Sandy Bridge-EX" variant as it had planned to do in 2012. Intel was working on a common socket for Itanium and Xeon E7 processors, then backed off on that.

The important thing is that Intel is moving from the 32 nanometer processes used to etch the Xeon E7 v1 chips straight to the latest 22 nanometer processes for the Xeon E7 v2 processors. By doing so, Intel can boost the core count on the chips by 50 percent, from ten cores to fifteen. This is a big jump in performance, and the main memory controllers on the new Xeon E7 v2 processors are able to address, at 1.5 TB per socket, three times as much memory as their v1 predecessors. This extra memory capacity may turn out to be far more important for a lot of workloads than the big jump in core count or the modest increase in clock speeds.

The other thing to consider, says Frank Jensen, performance marketing engineer in the Datacenter Group, is that the average refresh cycle for a server with four or more sockets tends to be five to six years, compared to three or four years for the typical two-socket machine. There are plenty of multi-processor machines out there with Xeon 7400s and 7500s as well as Opterons, Powers, Sparcs, and Itaniums that need to be upgraded or replaced. Many of those customers will stay on their current chip architectures, but others will see the benefits of the Xeon E7 v2 and make the jump to Linux or Windows. (Solaris shops will be able to stick with their operating system if they move to Xeon E7s and get rid of Sparc platforms as they see fit.)

With the large core count and even larger memory footprint, the Xeon E7 v2 processors are aimed at a number of workloads.

Obviously, any dataset that can be fit into a 6 TB footprint on a four-socket machine or a 12 TB footprint on an eight-socket machine that is particularly sensitive to network latencies is going to run well on the NUMA clustering infrastructure compared to a cluster using Ethernet or InfiniBand links with the same number of sockets and aggregate memory capacity. Examples include portfolio analysis in the financial services market or cosmology simulations in the astronomy field. The Xeon E7 v2 processors also sport large L3 caches and lots of threads, something that database management systems love. Companies that want to cram as many virtual machines on a node as they can – and yet have room to create very fat VMs on the machine when needed – will similarly take a shining to the Xeon E7 v2s. Service providers in China, for instance, tend to buy four-socket machines and only half populate then with processors, allowing them to add more oomph down the road without taking the machines out of service.

The Xeon E7 v2 processor has 4.3 billion transistors, and it plugs into the LGA2011 Socket R1 socket on server motherboards. Intel has not said so, but presumably this same socket will be used by the future EX variants in the "Haswell" family of chips implemented in 22 nanometer processors. It seems likely that the shift to 14 nanometer processes with the 14 nanometer "Broadwell" generation will compel Intel to make a new E7 socket, quite likely one with more physical memory attached to it as well as other changes. But maybe not. We shall see.

Next, the feeds and speeds. . . 

EnterpriseAI