Calxeda, a startup developing server-class system-on-chips (SoCs) with ARM cores, has demonstrated power efficiency of web servers based on ARM architecture. Apparently, a machine running on Calxeda's SoC is theoretically fifteen times more energy efficient than a comparable system based on Intel Xeon processor with normal voltage.
Calxeda used similarly configured systems with Ubuntu Server v12.04 (3.2 kernel) and Apache Server v2.4.2 to run ApacheBench v2.3 (16KB request size). Calxeda web server was based on EnergyCore ECX-1000 (four Cortex-A9 cores at 1.1GHz) and featured 4GB DDR3L at 1066MHz, 1Gb Ethernet port and one 250GB 7200rpm Serial ATA hard drive. Intel web-server sported Xeon E3-1240 (four cores with HT at 3.30GHz, 80W TDP) microprocessor, Intel C216 core-logic and 16GB of DDR3 memory.
Power measurements were obtained at a 2-second sampling rate and averaged across the duration of the benchmark run. Power supply overhead and hard drive power consumption were not included in these measurements, however the entire Calxeda EnergyCore ECX-1000 SoC and DDR3 memory are included together.
Based on benchmark results provided by Calxeda, Intel Xeon E3-1240-powered web server processed 26% more requests per second than Calxeda EnergyCore ECX-1000-powered machine and provided lower latency. Nonetheless, average power consumption of ARM server was 5.26W, while Intel web server utilized whopping 102W.
The power consumed by the Calxeda EnergyCore processor is actual measured power directly against real hardware. The Intel Xeon "Sandy Bridge" platform is based on published TDP values for the CPU and I/O chipset, along with an estimate for DDR memory since at the time of this testing, Calxed did not "have a way to measure actual power consumption with the same level of fine detail".
While the results clearly show advantages of Calxeda EnergyCore ECX-1000, the configuration of Intel Xeon-based testbed was not the most power efficient possible. There are quad-core Intel Xeon E3-1200-series chips with 45W TDP, besides, the system could have been configured with 4GB of memory instead of 16GB. Still, even with Xeon E3-1265LV2 (four cores with HT at 2.50GHz with 45W TDP) chip and 4GB of DDR3 memory the system would have consumed around 60W, considerably more than the Calxeda system.
Tags: Calxeda, ARM, Intel, Xeon, Sandy Bridge, Cortex
Comments currently: 8
Discussion started: 06/21/12 01:23:44 PM
Latest comment: 06/27/12 01:44:51 AM
Expand all threads
| Collapse all threads
Something here doesn't make sense.
First you say that the Xeon actually is 26% faster. Then you go on to say they used maximum power consumption (TDP) on the intel chip and compared it to actual power consumption on the ARM SoC.
So over 1 million requests. How could this chip be consuming 103 watts ? That's at TDP. It would be far fetched to believe the Intel chip to be running at 100% the whole time. You can't compare actual values on one, against max on the other.
Furthermore, the Intel chip is running at 3 times the clock speed and features more advanced execution units. I'd like to know how quad-core ARM SoC was able to supposedly match the performance of the Quad-core Intel chip. The time to process 1 million requests is 182 seconds on the ARM vs 144 seconds on Intel. Seems a little weak for Intel.
According to wiki. The DMIPS on a Cortex A9 Quad Core at 2.5 Ghz (this is a A9 at 1.1 GHZ, so substantially faster) is 15 DMIPS. The Xeon would be around 38 DMIPS(2600K 3.4 ghz vs 3.3 ghz). Per core that works out to 3.5 vs 9.43. So this SoC should be 1/3 the performance, if you removed the bottlenecks.
How could this thing possibly match the Intel chip in performance ? I would imagine the bottle necks held back the Intel CPU somehow. But even then, it wouldn't be running at MAX TDP the whole time. I would imagine that on a virtualized server their would be plenty to keep the Intel chip busy, even with the botttlenecks. Whereas this SoC would only be good for lightly used servers.
This almost seems like best case scenario for ARM vs absolute worse case scenario for Intel. Good marketing.
06/21/12 01:23:44 PM]
- collapse thread
Calxeda did test performance of both setups, but measured actual power consumption only in case of its own setup. For Intel, they published official max TDP values and did not count in power consumption of HDD.
Obviously, Xeon is faster. But I believe that the Xeon E3-1265LV2 will be at least as fast in terms of simple requests. Perhaps, a chip with even lower TDP can do the job...
The thing Calxeda wanted to say in that in many cases Xeon cores are too powerful for simple tasks like web servers and that compute power is wasted as the CPUs are I/O limited and hence consume power without making any job.
06/22/12 02:05:28 AM]
Mostly agree, but noticed mistake in your presumption concerning raw performance. Cortex-A9 architecture is said to have 2,5 DMIPS/MHz/core, so in theory:
Calxeda ECX-1000 @1.1 GHz ~ 11 DMIPS
Still much lower than Xeon. IMHO Apache in this test don't effectively use all Xeon internal CPU resources, so DMIPS are not relevant.
06/27/12 01:44:51 AM]
The biggest thing that was awful about this article is X-bit's own analyzing. You guys can do better than that, especially with the trouble you guys went through in creating a PCIe passthrough slot to measure power at the motherboard.
Still, even with Xeon E3-1265LV2 (four cores with HT at 2.50GHz with 45W TDP) chip and 4GB of DDR3 memory the system would have consumed around 60W, considerably more than the Calxeda system.
At max load, perhaps. Not necessarily when processing those requests. Intel's power management is among the best in the industry.
06/22/12 12:05:25 AM]
power management is it the most critical objective for most IT departments? most i´ve seen are services and deliveries of information and applications as needed by the company´s employees. when the numbers of systems approaching hundreds of servers in the data center than it´s time to look for raised-floor data center and air conditioner to cool down the servers. that´s when power management comes into the picture.
at the point when power management comes into the picture does not mean services and deliveries are eliminated.
what is the components that the IT could function to offer the services and delivery the information as needed is more critical.
06/22/12 01:01:38 AM]
Guy stop racking your brains, Intel Haswell will take care of everything. Haswell based server chips will be super energy efficient!
06/22/12 05:55:57 PM]
This test is just seriously stupid. Who compares exact measurements with maximum average numbers? The answer is: Calxeda's marketing.
Intel could easily answer with a web-benchmark using secure http / https. The ARM cores would loose big times (more than 15x), because they do not have on-chip cryptography.
My conclusion from this unnecessary marketing press release: Caldexa's product must be very bad, if they have to rely on this kind of biased testing.
Nobody needs apples to oranges comparisons.
06/23/12 11:01:30 PM]
look at the numbers: the xeon is saturating the 1Gb link, so the test is basically uninformative. it is slightly interesting that the arm can't quite saturate the link - interesting because the load is the dumbest possible unrealistic one: 16K static content. even a comparison using memcached would be more interesting, since such a workload would involve modest index/lookup work, as well as requests distributed over a larger-than-16k dataset...
in short: data does not support conclusions.
06/25/12 12:25:47 PM]
Add your Comment
Enter your username and e-mail address. Password will be sent to you.