By taking another step on Moore’s Law, using a low-voltage processor and changing to 10 gBE NICs, Intel was able to stay competitive with Calxeda in throughput per watt… proving what we knew all along, that ARM has a place in the server room:
“Without the overhead of the power supply (which only runs at 90 per cent efficiency or so), the fans in the chassis to keep the node cool, and the disk drive, the EXC-1000 microserver did 5,500 requests per second at 5.26 watts measured at the wall, which works out to 1,046 requests per second, per watt. That’s about 40 per cent better performance per watt against the Xeon E3 v2 with a 10GE port and about five times better than the Xeon E3 v2 with only a Gigabit Ethernet port. But this comparison is still not exactly fair, says Hill.”
Move over Intel, ARM is invading your spaces this year.
see Intel: Xeon breaks Calxeda's ARM in Apache benchmark • The Register.
17460
12760
206
3
2
23903 
11858 
11710 
4633 
4259 
1641 
198 
14 
2 
0 
0 
0 
Mr. Pogson, this is getting sad.
Do you even try to read the articles you link to these days?
The article you linked to is about how Calxeda rather stupidly tuned the benchmark to its advantage, and the article’s conclusion is that Calxeda’s ARM setup didn’t offer a better per-watt-performance than the Intel setup.
The “Cult of ARM” tries to strike again. But like a poor marksman it constantly misses its target.
Chris Weig, completely missing the point, wrote, “Calxeda’s ARM setup didn’t offer a better per-watt-performance than the Intel setup.”
The point is that even without the latest step in Moore’s Law, ARM was competitive. Of course, people who think Wintel is wonderful don’t get that other technology can be competitive, but let me explain. When ARM is out on sub-20nm and Intel is as well, the power issue will drop off the charts and throughput and cost will dominate the discussion. ARM has been proven to be competitive. That changes everything because ARM is much cheaper. That one of the first solutions to arise using ARM can do that is earth-shaking. The world is noticing even if Chris refuses to pull his head out to look.
“ARM is much cheaper”
Really? OK, let’s do some math here – I know you love that.
A professional server (rack-mounted, with redundancy all over the place, remote-admin board etc.) starts at ~ $3,000. Let’s be generous and say that you can save $400 by chosing an ARM CPU and the slower NIC. Now look at the performance figures from the article you quoted:
ARM: ~ $2,600 5,500 requests per second
Intel: ~ $3,000 35,624 requests per second
Which system has a better price/performance ratio?
The problem here is that in absolute terms, ARM is cheaper and consumes less power exactly because they have such a low performance. And with the quoted Intel CPU costing only $189 but delivering more than six times the performance, things get tough for ARM.
ch wrote, “ARM: ~ $2,600 5,500 requests per second
Intel: ~ $3,000 35,624 requests per second”
There is no Intel product for $3000. You are comparing one module of ARM with a different package entirely from Intel. Compare the full box:
“A 2U rack enclosure could hold 120 server nodes: that’s 480 cores.” see http://www.theregister.co.uk/2011/10/26/hp_calxeda/
How much throughput can Xeon put into 2U and at what price?
What’s the point of comparing four cores of ARM with one Xeon? Compare 480 with 48 … Oops. In the big picture ARM is very competitive. If one 4-core socket threatens Intel enough to benchmark it. Imagine what 10 would do. ARM is competitive even at 56nm. Imagine how good things will be at 20nm. Intel is on borrowed time. ARM is the new commodity server chip.
Boston has 192 ARMed cores in 2U at 300W. That’s competitive. They are shipping.
Pogson, how many times to I have to tell you and bring proof that more cores do not result in higher productivity?
Gosh, can’t you learn anything from AMD?
““A 2U rack enclosure could hold 120 server nodes: that’s 480 cores.””
Mr Pogson, I was simply refering to the benchmark in the artivle you quoted: It involved one quad-core CPU on either side, so what do 480-core machines have to do with anything?
Besides, if this benchmark is any yardstick then 480 ARM cores would be roughly the equivalent of ~60 Intel cores or eight eight-core (octa-core ?) CPUs.
ch
“And with the quoted Intel CPU costing only $189 but delivering more than six times the performance, things get tough for ARM.”
Ok this is the problems.
“remote-admin board” That is in Calxeda card. The card that holds 4 Calxeda processors so 16 cores. The card that costs 1/4 of the intel price.
This here you made mistake. “ARM: ~ $2,600 5,500 requests per second” You have to times the request by 4. Since you don’t by the arm cpu solo. 22000 requests per single card.
480 ARM cores is 30 cards. You get 12 cards per 2U with arm at light density allowing for hard-drives per card. So 6U of rack space. Remember in 6 U you have fitted as well over 30 hard-drives.
Or high density Half a 2U will hold 480 arm cores or basically 1U of space. I don’t see you fitting 8×8 cores x86 in a 1U case or half a 2U case and not having heating problems. Mind you even at high density you can fit a lot of storage.
“MicroSD Card support via four (4) slots” basically MicroSD Card per Processor. 64GB max basically. 30x4x64 or nice 7 TB of storage solid state evenly spread between your cpus. Or 256GB of storage each card you insert.
That does not allow for the 16 sata ports the card has. Yes low density 12 per U2 to fit the all harddrives the arm boards support requires using 2.5 inch hard-drives. 12x16x500GB to be nice or about 95 TB of storage. Yes the hard bit is fitting the 192 hard drives. Most of the light density cases is hard-drive storage with the Calxeda. So its a full length rack-case and heavy as hell when full.
ch the density of arm is basically insane. The harddrives connected to the card generate more heat than the card. So might manage compare x86 to a arm in low cpu density packing. Problem is x86 does not have direct hard-drives connected in that or the am-mount of decanted flash storage per cpu.
So ARM either lowish cpu and insane storage or high cpu and reasonable storage with arm. Even the lowish ARM cpu count per box x86 has trouble matching. Mostly because the x86 motherboards are too big for what they are.
The other evilish thing about this system is where you don’t fill with cpu cards you can fill with hard-drives.
ch x86 simply does not have the density yet. Arm 64 bit will make the problem worse.
Phenom
“Pogson, how many times to I have to tell you and bring proof that more cores do not result in higher productivity?”
True you have to bring more memory controllers as well and more network cards …..
Otherwise you bottleneck.