Microsoft Pressing Intel for 16-Core Atom
A 16-core Atom in Microsoft's servers could possibly maintain the current level of performance but consume less energy.
Microsoft is reportedly pushing Intel to create a 16-core version of its low-power Atom chip. The reason behind the request makes perfect sense: to use these processors to reduce the overall amount of power consumed by servers stationed in massive data centers--especially those powering Bing, Hotmail and Windows Live Messenger-- without sacrificing performance.
In a speech presented at The Linley Group Data Center Conference in Silicon Valley, Dileep Bhandarkar, an engineer with Microsoft's Global Foundation Services, said that there's a "huge opportunity" to improve energy efficiency by using servers based on processors like Intel's Atom, AMD's Bobcat or Via's Nano.
As it stands now, Intel processors-- like the Xeon--crammed in Microsoft's servers bring high clock speeds but are burdened with heavy power consumption and active cooling, thus costing the company a pretty penny. By using a 16-core Atom processor, Microsoft's thinking is that it can achieve the same high-GHz performance using the low power multi-core chips, thus requiring less power and cooling. Using a system-on-a-chip design would be even better.
"When you look at these tiny cores, another way of making them work in a very efficient way is [not to] surround them with a whole bunch of south bridges and network controllers," he said. "Essentially, the tiny cores and systems-on-chip should go together."
Bhandarkar expects Intel to eventually cave in now that ARM is heading into the server sector, adding that Intel will need to rely on the Atom in order to compete with ARM on power performance per watt. That said, there's a possibility Microsoft could incorporate ARM-based servers in the future. The only drawback is that the architecture will need to overcome a few serious hurdles first while also showing a clear performance benefit over x86 solutions.
"Instruction-set transitions are extremely painful," Bhandarkar said. "As a general rule of thumb, you have to have a sustainable improvement per dollar per watt of at least 2x -- some would say 5x -- but it's at least 2x to make it worthwhile. For some apps where you don't have that dependency the number could be smaller. ARM's an interesting thing to look at and, if nothing else, if it lights a fire under Intel and AMD to deliver more effective x86 solutions, I'm happy."
Currently Intel has no announcements to make in regards to Atom chips for data centers.
That is what they want because the Terascals is highly customizable unlike Atom.
intel runs the world, get used to it
correct me if im wrong, but isnt server side of things not dependent on speed but on how much data it can process? the 16 core may = the current ones in data processing power, but would use what is it, 1/5 the power?
arent those experimental and no where near mass production?
because code would need to be rewritten to take advantage of an apu, and that is a massive ammount of work, on par with porting it to arm.
Server code can't really run on the GPU part of a APU, at least not web/file server code. The GPU part of a APU can only run massively parallel floating point code, web/file servers don't use much if any floating point math so the APU would be of no help.
I could see a atom with 16 cores doing well in a cloud computing server. A major tenet of cloud computing is to use low power/low cost CPUs in large numbers and so a low power 16 core atom would be a good fit.
http://www.tilera.com/products/processors/TILE-Gx_Family
lets put this in a way they people can understand better. a watt per core.
xenon 3.33 (3500-series "Bloomfield") would be at best, 16.25 watt per core, and at worst 32.5 watt per core, depending on if you count logical as being as good as a physical, i dont know how that works server side, but im assuming its better than desktop.
the best xeon, (6500/7500-series "Beckton") which i doubt they use, comes to 16.25 watt per core, on physical, and 8.125 watt per core if you count logical.
now the atom side, i went with the top model, and the lowest power model.
the lowest power, daimondvill dual core ht, comes in at 4watt per core, or, 2 watt if you go logical
and the best one, pineview, which comes in at 6.25 watt, and 2.75 watt pr core.
now lets take a look here, at a 1000 cores and in best conditions (these servers i assume use it in the 10000+cores per service minimum, probably that amount of cpus) now servers on the scale of hotmail and the like dont upgrade often, at least i dont think they would, but they make big upgrades, so Microsoft may be using even more inefficient processors.
but here, 1kwatt is 1000watt. and kwatt here will be measured in hours, and where i live 1kwatt equals 11 cents.
xeon 16.25kwatt = 1.7825$ an hour
xeon 8.125kwatt = .89375$ an hour
atom 2kwat = .22$ an hour
atom 2.75kwatt .3025$ an hour
lets blow this out, to 100000 cores to give you a better picture.
xeon = 178.25$ an hour 4,278$ a day 132,618$ a month 1,561,470$ a year
xeon = 89.375$ an hour 2,145$ a day 66,495$ a month 782,925$ a year
atom = 22$ an hour 528$ a day 16,368$ a month 192,720$ a year
atom = 30.25$ an hour 726$ a day 22,506$ a month 264,990$ a year
and this is powering the cpu alone, and those numbers are probably higher than i just mentioned. i just expanded it to the point people can see the real world difference
i should also point out that arm apparently has tapped out a 28nm chip, cabable of up to 2.8ghz, and current arm cpus use less than 4 watts total, some older ones useing less than a watt, so i cant say how many watts they can put out, but what i do know is this, an arm chip can be put in a ipod, or other devices, and make nearly no heat, compared to competitors, so putting lets say 20 or so of those on a motherboard (if possible, o know we have 4 cpu mbs right now, but i think they dont have more due to size constraints on the cooling they need, and arm are SIGNIFICANTLY smaller than its competitors.
yes and no. yes in the sense that its more than fast enough, and no, in the sense that the code wouldn't execute off that chip
Yes Bobcat is faster than atom on the CPU side according to passmark, just compare the scores of the e-350 to the fastest desktop Atoms
E- 350 gets a 798
Atom D525 gets a 718
http://www.cpubenchmark.net/cpu_list.php