Stacking Up Xeon E7 v2 Chips Against The Competition
Since Intel entered the server market formally in 1993, it has used its prowess in manufacturing and its ever-expanding sophistication in chip design to knock countless competitive chips out of the market. The success of Intel’s Xeon strategy cannot be denied, especially when you consider that Intel has been able to turn its staunchest rivals in the datacenter into its tightest partners.
With the new “Ivy Bridge-EX” Xeon E7 v2 processors, aimed at the high-end of the server market, Intel is going to be able to put even more pressure on the remaining makers of non-Xeon systems. This group includes IBM with its Power7+ and System z machines, Oracle with its Sparc T5 and M6 systems, and Fujitsu with its Sparc64 machines, to name the ones that still have a sizeable revenue stream. It also includes Hewlett-Packard, Bull, NEC, and a few other makers of Itanium-based systems, by the way, even though Intel did not bring that up at the Xeon E7 v2 product launch. The company’s competitive analysis focused mostly on Power and Sparc machines, with a comparison or two with AMD’s Opterons, just for good measure.
In our analysis of the Xeon E7 v2 processors on launch day, we already went through the feeds and speeds of the chips and detailed price/performance analysis of the new Xeon E7s versus the prior generation. We also looked at how the Xeon E7 v2 chips compare to the Xeon E5 v2 chips aimed at two-socket machines, which debuted last September. In this story, we will look at how four-socket machines based on Power7+, Sparc T5, and Opteron 6300 processors compare to those using the new Xeon E7 v2 chips.
But just for fun, here is a chart that Diane Bryant, general manager of the Datacenter and Connected Systems Group at Intel, showed at the launch event in San Francisco.
It compares the transaction rates of a financial trading system for a Sun Fire E25K system from Sun Microsystems, circa 2004, and compares it to a quad-socket machine using the Ivy Bridge-EX processors from Intel. The Sun box was top-of-the line at the time and ran Solaris Unix, which was very popular in the financial services market from the mid-1990s to the mid-2000s; it was able to handle about 26,000 messages per second on an unknown financial workload (garnered from industry experts at TABB Group). The precise configuration of this machine was not divulged, but Intel said it cost $775,936. A four-socket server (from an unspecific vendor) using four of the top-speed E7-4890 v2 processors was able to handle 6.5 million messages per second, all for a mere $51,237. (That system was configured with 256 GB of main memory and two disk drives.) That’s a factor of 250 improvement in messaging throughput, with a factor of 15 reduction in cost. To put this in more concrete terms, the base Sun Fire E25K system cost nearly $3,000 per message per second, while the current Xeon E7 v2 machine costs eight-tenths of a penny per message per second.
In industry after industry, through hundreds of thousands of commercial applications and untold numbers of homegrown applications, has come to dominate just about all of the server workloads. This has significantly reduced the profits from hardware in the high-end system business, which is not necessarily good for the IT industry even if it is good for customers. This sharply downward pointing price/performance curve is precisely why so many companies have shifted from proprietary and Unix platforms to X86 machines running Linux and Windows. It is also the kind of curve that emboldens an army of ARM server chip makers, because they think they can bend it more.
As the dominant maker of Unix machines, stacking up the new Xeon E7 chips against the latest Power7+ boxes from IBM is the first place Intel started:
This comparison pits a four-socket Power 750 express server using eight-core Power7+ chips running at 4.0 GHz against the fifteen-core E7-4890 running at 2.8 GHz. The IBM machine tops out at 1 TB of main memory, compared to 6 TB for the new Xeon E7. IBM can deliver four threads per core, and Intel can do two per core. But as you can see, it does not all even out. Depending on the application type, the Xeon E7 server enjoys a 40 to 82 percent performance advantage over this IBM machine.
For memory intensive workloads and database processing, the quad-socket Sparc T5-4 system is more competitive with the Xeon box, by Intel’s internal analysis, and in general, it has closer to the same performance as the Xeon E7 v2:
And the Sparc T5-4 machine doesn’t do so badly on HPC workloads either, but Intel still beats it on everything but the memory bandwidth benchmark. The Sparc T5 processor has sixteen cores and eight threads per core; the cores run at 3.6 GHz, which is the only speed that Oracle offers.
For the Opteron comparison, Intel gathered up stats for what it called the “Opteron 6300 baseline,” and we will presume that Intel was talking about a quad-socket server using a sixteen-core Opteron 6380, which has a clock speed of 2.5 GHz and which is the fastest standard part offered by AMD:
As you can see, the Xeon E7-4890 v2 offers a significant performance advantage, and a lot of that can be attributed to the architecture of the processors since the two chips have clocks running at nearly the same speed (2.5 GHz and 2.8 GHz) and nearly the same number of cores (sixteen versus fifteen). Intel is basically getting twice as much work done per socket.
AMD has said very little about its plans to enhance the “legacy Opterons,” as it now calls them in the wake of its impending delivery of the “Seattle” Opteron A1100 64-bit ARM processor. In January, AMD delivered slower and more energy efficient versions of the Opteron 6300s, code-named “Warsaw” and aimed specifically at financial services and cloud customers who asked for these components. While these offered better thermals and better bang for the buck, Intel is examining top-end performance only in its comparisons. AMD could still have a thermal advantage in configured systems. We shall see.
The most interesting chart Intel created showed the relative value of bare-bones systems on the SPEC integer benchmark:
To be precise, Intel dug up performance numbers for the Power 750 and Sparc T5-4 quad socket machines on the SPECint_rate_base2006 test.
The Xeon system in the comparison above had a score of 2,238 on that test and cost $46,500 in a configuration with 256 GB of memory and two disk drives. That works out to $20.77 per unit of integer processing on the SPEC test. This machine was configured with Red Hat Enterprise Linux and the integrated KVM hypervisor.
The Power 750 machine had 4 GHz Power7+ chips with eight cores per die, plus 256 GB of memory and two disks; it was set up with IBM’s AIX Unix variant and its PowerVM hypervisor, and all told, it cost $177,290. It was rated at 1,230 on the SPEC integer test, and that works out to $144.14 per unit of performance. That is a factor of 6.9 higher cost per unit of work. Even when IBM delivers Power8 chips with twelve cores and more than doubles the performance on integer work per socket at the same 4 GHz clock speed, as it has said it can do, the value gap will still be significant if IBM does not bring the cost down and the clock speeds up. It can cut the price with PowerLinux configurations, which have lower hardware prices and also are configured with cheaper Linux licenses from Red Hat. IBM now knows exactly what target it has to hit.
So does Oracle if it wants to compete with the Xeon E7 v2 systems with its Sparc T5-5 machines. The Oracle quad-socket server was configured with 1 TB of main memory, so Intel mirrored that in the chart above and the Sparc machine cost $147,992 with Solaris on it compared to the Xeon E7 machine with REL at $57,729. The Sparc T5-4 server was rated at 1,745 on the SPEC integer test, and that works out to a more than three times higher price per unit of work. If Oracle charged the same price as Intel for its server, it could close the price/performance. But doing so will sacrifice a lot of profits – something Oracle co-founder Larry Ellison did not buy Sun Microsystems four years ago to do.
Oracle has not said much about the future Sparc T6 chip, but in late 2014 or early 2015, this chip could come out and will have a 50 percent increase in per-thread performance (a mixture of clock speed and microarchitecture improvements) and twice as much overall throughput at the socket level. So Oracle might be able to close the gap early next year and surpass the Xeon E7 v2 on integer performance by 50 percent or so. But that will be just in time for Intel to start its next run with the “Haswell-EX” Xeon E7 v3 chips.
There’s no rest in the server chip biz. Wicked or not does not matter one bit.