Slashdot Mirror
AMD Beats Intel in Power-Efficiency Study
Ted Samson writes "AMD Opteron servers proved up to 15.2 percent more energy-efficient than those running Intel Xeon in a server-power-efficiency test performed by Neal Nelson and Associates, InfoWorld reports. That translates to annual electricity savings between $20.29 per server and $36.04 per server, depending on the workload, the study concluded. The benchmark tests were conducted on similarly configured 3GHz systems running Novell SUSE Linux, Apache2, and MySQL."
I would have liked to seen them test it with the 'big three' OSes of Linux (RH and SUSE), VMware and WIndows. It would have been nice to see if the power management of the operating systems would have come into play some above and beyond just the single OS. Besides the OS the applications used run on any of those platforms.
As a rock-in-roll Physicist once said, No matter where you go, there you are.
I run AMD because I own their stock because I run AMD because I...
Did AMD sponsor this, by any chance?
Over what time period? Or are they using the prime interest rate to figure out what the one-time savings are?
Open source software tends to bring reduce power consumption, as well. After switching a number of our systems from Windows Server 2003 to Linux, we saw a fairly significant drop in our electricity costs.
Our analysis suggests that this was due to the open source software being more efficient than the equivalent Windows-based software. This is backed up by the fact that we saw a significant performance boost after the transition. Database jobs that would take 20 minutes on SQL Server 2005 and Windows 2003 would be done in under 5 minutes when using PostgreSQL on Linux. Those 15 minutes of not processing resulted in far less usage of the CPU, and far less disk activity, thus saving power.
If you read the PDF, you'll see that the AMD system was tested with a 500W power supply while the Intel one was tested with a 600W one. I wonder how much of the different can be associated with that.
That translates to annual electricity savings between $20.29 per server and $36.04 per server
Did you seriously not read up to the fourth word in that sentence?
The article did not say "$20.29 per server and $36.04 per server annually" but said "annual electricity savings". That means, if a server would have used 100 Watts but now uses 50 Watts, that's an annual saving of 50 Watt-years. However, that doesn't tell us how LONG it takes you to save fifty watt-years every year. It's a valid question, since it's not like the savings is instantaneous - you don't just click a button, and on the up-click BAM the savings are realized. It takes time and equipment to realize the savings and the new equipment doesn't last forever - it's not like you can just make the savings once and then keep reaping 50 watt*years less per year forever. So the parent poster has a good point, even if its hard to see it.
A 3.0GHz Core 2 is more than 15% faster than a 3GHz Opteron for many tasks.
AMD is doing better at idle speeds (Intel definitely needs to crank Penryn down more when it's not in user) but if this survey compared equivalent performance processors, the difference would be much smaller.
Both systems had 3.0Ghz CPUs and similar amounts of RAM. But did they offer the same performance? If both servers were being pushed 100% would one be able to server more users than the other? If the servers were never pushed to 100% then the test is not really a like-with-like comparison. I imagine that one CPU performs better than the other (and I'd expect right now that's the Intel one). Perhaps a 2.66Ghz vs 3.0Ghz test is closer to the same performance?
I just bought my second AMD system. Both require that I enter a NOAPIC on them because they have such squirrelly set-ups. That means that I lose the advantage of having any power control. Since AMD does not push their MOBOs to work right, tt will be my last AMD that I buy.
If add that ram for AMD severs costs less then FB-DIMMS.
Also FB-DIMMS and the intel chipset need a lot more power then amd chipsets and DDR2 ECC / DDR1 ECC ram.
If you look at the systems a bit more, you'll see that Intel's requires a bigger power supply because it requires more power.
AMD 8222 cpu = $2149
Intel 5160 cpu = $851
The AMD system will be obsolete before you realize any "cost savings".
Also you don't buy these top dog chips if you're going to let them sit idle all day.
We see similar when we build systems.
The Intel CPUs are competitive with the Opterons on power consumption.
But: The whole system uses more with Intel.
Why? the northbridge memory controller is a separate chip with Intel, and it is very power hungry.
In the AMD chips the memory controller is a part of the CPU.
In the case of a similar dual XEON compared to a dual Opteron,
the XEON machine uses about 80W more power.
What a lot of these studies do not even get into is cooling cost.
for every watt of power , which ends up heat, we have to expend at least 1.5 watts, on air conditioning.
As for the comment about the size of the power supplies, that is irrelevant.
The maximum rated output of a supply has nothing to do with the power consumed.
Bottom line:
Assuming an Intel XEON server uses about 80 watts more than an equivalent AMD one,
which is what we see when we build them:
80w x 24 hours/day x 365 days is 700KWh. @ 9c/kWh costs $63/year.
Add aircon costs for that extra 80W:
120w x 24 hours/day x 365 days is 1050KWh. @ 9c/kWh costs $96/year.
Therefore, a machine using an extra 80W costs an extra $160 to run in an air conditioned room.
Source of power rates:
http://www.neo.ne.gov/statshtml/115.htm
Maurice W. Hilarius Voice: (778) 347-9907
Render farms, HPC, etc. are a tiny percentage of all the "servers" out there.
I don't think we should throw away the test results because of a few render farms.
No sig today...
There is so much bullshit like this out there. All these alleged experts keep telling me how cheap hardware is but none of them EVER seem to be able to put together two identical machines for a bake-off. They're not just slightly different, as might be required by something as substantial as differing CPUs, they're wildly different.
pseudo-science. They might as well be killing chickens.
Unless these computers are similar in performance and modernness, it makes no sense to compare them for power consumption.
The woodcrest platform used for the Intel side of this test was a year old, while the AMD platform was their latest and greatest. This kind of test is crap until they test the latest and greatest AMD latest and greatest Intel - same power supplies, etc. Of course, to be valid at any given time, this requires that these kinds of tests be run after each release of a new platform by either company. This is a result only an AMD fanboy could love.
What kind of crappy A/C do you have? I would expect more like 0.5 watts for air conditioning.
Slashdot gets worse every day... Pipedot: News for nerds, without the corporate slant
Your idea is right, but your math is a little off.
You should be able to get down below 1.5 kW per Ton of A/C. (efficient systems can get down below 1.0 kW/T, even including all the pumps and fans)
That works out to close to 0.4 kW of A/C power used per 1.0 kW of heat cooled. But first add about 0.15 kW UPS per 1.0 kW power delivered, so you might as much as 0.5 kW per 1.0 kW of server power.
The maximum rated power supply does not correlate to power consumed, but an over-sized or under-sized power supply will be less efficient.
Also, is that 80W power differential the average over 24hrs/7days a week, or is it closer to a peak use difference?
Did I get something wrong or just calculate with horribly over-efficient air conditioning?
Actually I was agreeing with you. If you read what wrote I said that was correct for most servers. However with XEN I wonder how the cpu loads will lock in the future.
See my blog http://ilovecookes.blogspot.com/ for light hearted technical information.
Now, that would be interesting.
But since Infoworld is more accurately called "Wintelworld", they'd never publish that.
I have not seen any clusters that achieve consistently high utilization (though this is often a goal and sometimes inflated by whoever is backing the cluster to prove its worth). I don't have direct experience with render farms, but I'd imagine it's the same; there are "crunch times" when they're saturated for a few days (or up to a few months) on end, but then that project ends and there's less load for a while. Anyways I agree efficiency under load is important, too.
There are so many questions to answer for this question that the real and true answer will never be known.
AMD runs better with no load, where Intel runs better with full load. So in this particular instance, do you have a server that's gonna idle 99% of the time(If so why are you not using a VMWare setup?) I'd expect a nice new server to be cranking out 100% usage for as long as I can keep it there.
You are correct that for every 1 watt of heat, it takes 1.5 watts(or sometimes even more) to remove the heat.... BUT the power supply is also relevant for 3 reasons:
1 The power supply does 'expend' energy that is basically lost(as heat).
2. This heat must be factored into the heat generation if you are taking into account the cooling.
3. Every power supply has it's own power efficiency curve. And size does matter(ladies?). I'll use 2 Seasonic's as examples because I own one myself:
-500W http://www.80plus.org/manu/psu/documents/Seasonic_ SS-500ET_500W_80PLUS_Report.pdf
-650W http://www.80plus.org/manu/psu/documents/Seasonic_ SS-650HT_650W_80PLUS_Report.pdf
Notice the most efficient curve peaks at roughly 50%. That means that if you are looking for pure efficiency, you need to know how much power your machine uses(at the outlet of the power supply) to determine how to make your system as efficient as possible. The 15-20% loss is mostly I(squared)R losses(heat loss). So you see size does matter. It may matter for 13% heat at 250W or 19% heat at 600W, but it DOES matter. The only true way to compare is to show watts consumed at the output of your power supply, not the input. If I were building a system for efficiency, I'll match my own power supply to my loading requirements.
The other thing to check out is performance and component cost. If your server is gonna be run at 100%(ideally your IT budget prefers this because you're maximizing your $), the Intel may be able to service more 'customers' at the same time than AMD.
Here's a hypothetical scenario:
I'm setting up project D. Project D will have to supply 10,000 customers simultaneously 24x7. I'm a big company and those server will actually have 10,000 customers serviced 24x7(this is only to provide static loading).
If I buy Intel servers I'll need 100 machines.
If I buy AMD servers I'll need 200 machines.
So if my Intel servers use 80W more power than AMD:
Intel Watts Total = 100 machines x watts used
AMD Watts Total = 200 machines x watts used
You could also say: Intel Watts Total = 100 machines x (AMD watts for 1 machine + 80W)
So if you need 100 Intel servers that draw an extra 80W of power each to service your 10,000 customers, but you need 200 AMD servers to service the same customers, which becomes more efficient? Suddenly your buying twice as many machines(so posibly higher startup costs for AMD), and I'm sure each of those AMD machines will draw more than 80W for each machine(remember the different between the AMD and Intel in my scenario is Intel is 2x more powerful for only 80W, so 2 AMD machines should be able to do the same work for cost of 1xAMD Watts + 80W to stay even with Intel for efficiency). But now you have twice as many systems that can fail, they are all highly loaded(which leans in intel's favor), more IT support will be required for those machines, more cooling for those extra machines, more space in a building somewhere(assuming you have the actual space), etc...
So, which setup was more efficient? In this case, Intel was. But if AMD's machines only drew 5 watts of power at full load, and their machines were 1/2 the cost, then maybe they'd be a better bet. I don't know about other people, but space in my building is at a premium. You can theoretically buy an infinite number of computers,
Surely $20 -$30 per year is a joke in a business environment, my office would save more money buying slightly cheaper coffee than by moving away from zeon's in our small datacenter.
I wonder... people get so hooked on Intel vs. AMD comparision... but what about thinking the total architecture?
Like, you start with software. They used Apache, Linux and MySQL -- what about, say, lighttpd, BSD and PotgreSQL? Each is reputedly more efficient than its counterpart. And what about comparing to a sytem with a different architectural decision, like business rules in the DBMS à la Alphora Dataphor (partly doable in PotgreSQL or IBM DB2) or like Lisp?
Moreover, one should compare to RISC too: Niagara and POWER6 come to mind, but here x86's economies of scale won't allow a level playing field.
Leandro Guimarães Faria Corcete DUTRA
DA, DBA, SysAdmin, Data Modeller
GNU Project, Debian GNU/Lin
With the A/C systems I've seen, I could have sworn that they run at 33% the power vs the heat they pump.
Since everybody knows that AMD outperforms Intel on a per-GHz basis, it does lead one to wonder why they chose that particular metric, but honestly no matter what metric they chose people would complain.
Is that still true considering that Woodcrest is based on the Core (not Netburst) architecture. Core should outperform Opteron per GHz quite a bit by now. Reading the comments on TFA confirms this without having to pull out some Anandtech benchmarks.
Where do yet get your 100*Intel = 200*AMD ?
Live long and prosper...
If this were true we'd all be heating our homes with such devices!
"The likes of Facebook and WhatsApp are free to those whose privacy is of zero value."
I didn't say it "destroys" three or four times the energy it consumes, I said it "pumps" that amount of energy from one end point to another.
If the temperature outside is higher than inside, using an air conditioner to heat your home makes perfect sense, but I don't think it's very probable that the desire to heat your home coincides with that condition.
I would think that render farms would be cranking out frames or would be in sleep mode with little to nothing in between.
Of course that is just the logical way they should work which means that it is probably wrong.
See my blog http://ilovecookes.blogspot.com/ for light hearted technical information.