However, at least one tech site is questioning the legitimacy of the results, saying the testing methodology was flawed.
Calxeda is
touting a benchmark test that shows one of its ARM-based EnergyCore chips being
15 times more power-efficient than an Intel Xeon E3 server chip, the latest
move in the growing competition between Intel and ARM.
The test
included a server powered by a quad-core EnergyCore ECX-1000 system-on-a-chip
(SoC)based on ARMs Cortex-A9 designwith a speed of 1.GHz and running Ubuntu
Server v12.04 and Apache Server v2.4.2. The results for the Intel chip were
based on published thermal design power (TDP) metrics for a quad-core 3.3GHz
Xeon E3-1240, which is based on the older Sandy Bridge architecture, and
estimated double data rate (DDR) memory.
The
results, posted to the
armservers.com
Website, indicated that with the EnergyCore system, for 1 million requests,
the server averaged 5,500 requests per second with a 9-millisecond latency and
average power of 5.26 watts. In the estimated Intel configuration, the results
were 6,950 requests per second, with a 7-millisecond latency and average power
of 102 watts.
The
results indicated that the EnergyCore SoC had a 15-times advantage over the
Intel chip in the performance-per-watt category, a key consideration for ARM
and its manufacturing partners.
The
results, coming amid the heightening competition between Intel and ARM in such
areas as mobile devices, PCs and servers, got the attention of some news sites
as well as Forrester analyst Richard Fichera, who has talked about the possible
benefits of the ARM architecture finding its way into data centers.
In a
blog
post June 20, Fichera noted that this was the first well-documented
benchmark test between an ARM-based server chip and one from Intel, and that it
is only a single test that is limited to the workloadWeb serving on a small
server with a 1G-bit network. The results cant be extrapolated to assume a
faster networkthough Fichera said he didnt think that would make much
differenceor other workloads.
Still, he
said, the test did show some interesting facts, including that the Calxeda chip
did demonstrate that an ARM SoC can deliver impressive power efficiency and
that the absolute performance was better than expected. As a basic proof
point, this benchmark succeeds as a proof of conceptARM servers are indeed in
the ballpark versus their initial promises, Fichera wrote.
Even by
swapping out the Sandy Bridge chip for a more power-efficient Ivy Bridge
processor from Intel, the performance-per-watt advantage for the ARM chip might
drop to 10 times, he said. Adding in the performance advantage of Ivy Bridge
over Sandy Bridge, that might further reduce the advantage to eight times.
Still
very compelling from an overall efficiency perspective, Fichera said.
Not
everyone is buying it.
On the
Website
Decrypted Tech, writer Sean
Kalinich said there are too many flaws in the test methodology to create
accurate results. In particular, Kalinich noted that the results for the Xeon
server were based on published TDP numbers for the CPU and I/O chipset, and the
DDR memory estimates. According to the results reported on the armserver.com
Website, unfortunately, at the time of this blog post, we didnt have a way to
measure actual power consumption with the same level of fine detail.
Now that
right there makes the whole claim invalid, Kalinich wrote. You cannot call it
a comparison if you do not have both in the lab.
He also
said that unlike in the test, most IT administrators would not use an entire
Xeon E3-1240-based server for a single Apache Web server. Instead, they would
leverage it as a host for 10 or more virtual servers, driving up the system
utilization numbers, which could offset advantages in total cost of ownership
claimed by the ARM-based server.
As Intel
makes an aggressive push into the mobile device market with chips aimed at
smartphones, tablets and Ultrabooks, ARM is looking to move into PCs and
low-power servers. While the companys ARMv8 architecturewhich boasts such
data center features as 64-bit capabilities, more memory and greater
virtualization supportis still at least a year away from appearing in systems,
ARM is still seeing some interest among server OEMs. In May, Dell announced
plans for limited distribution of its
new
Copper serverspowered by ARM-based chips from Marvellto help grow the
ecosystem around ARM-based servers.
That came
several months after Hewlett-Packard, in November 2011, announced a
deal
with Calxeda to develop very low-power servers as part of HPs larger
Project Moonshot initiative. However, HP officials said June 19 that the first
Gemini servers coming out of Project Moonshot would run on
Intels
Atom-based Centerton SoC platform.
Email
Print
