HP ditches ARM for Intel in Project Moonshot
Intel's Atom chips will now be at the heart of the first commercial product to appear from HP's Project Moonshot, in place of the originally-planned ARM processors.
Originally announced back in November 2011, Project Moonshot is HP's attempt to create low-power high-density server systems packing as many as 2800 server blades into a single rack. The company partnered with ARM, AMD, Calxeda, Canonical and Red Hat on the project - a list of companies in which Intel, you may note, does not appear.
The original plan was to create servers based around Calxeda's EnergyCore dual-core chips, which feature a class-leading thermal design profile (TDP) of just 1.5 watts. Using these low-power chips, HP claimed it could create servers drawing 89 per cent less energy and taking up just six per cent of the space of a traditional server installation. That equates to massive improvements in efficiency for high-volume datacentres and a dramatic saving in power and heating, ventilation and cooling (HVAC) costs for everyone else.
It also marked ARM's first major foray into the world of the server since the company departed the desktop market following the industry's switch away from ARM-based systems like the RiscPC and towards IBM-compatible x86 devices. For ARM, HP's decision to use its instruction set in Project Moonshot was a major validation and a win which could be used to great effect in convincing other high-density server makers to investigate ARM-based solutions.
That's a situation which Intel, clearly, has found untenable - and it has convinced HP that Atom is the way to go.
According to HP's latest announcement, the initial production servers to come out of Project Moonshot - codenamed Gemini - will be based around Intel Centerton Atom chips, rather than Calxeda's EnergyCore products. Intel, naturally, is saying that HP's made the right choice. 'The unprecedented value of the Intel Atom processor codenamed Centerton — with features like 64-bit support, ECC and an established software x86 ecosystem — will offer customers the increased flexibility and drastically reduced total cost of ownership required to compete in the fast-growing hyperscale computing space,' Intel's Jason Waxman, general manager of cloud infrastructure at the chip giant, claimed of HP's move.
HP's decision to support Intel despite the latter company's refusal to become an official partner in Project Moonshot is no real surprise: back in November HP announced that the second-generation Moonshot servers, codenamed Redstone, would include an Atom option - but it was always planning on using ARM chips for its first-generation Gemini product.
HP isn't giving up on ARM in the datacentre altogether, though: the company has confirmed that, while Gemini will launch later in the year using Atom chips, processors from other, unnamed vendors will be produced for use with the blade infrastructure. The company did not, however, provide a timescale for their release.
Previous Article
22 Comments
Discuss in the forums ReplyThe current one is shite compared to ARM's offerings.
Explain. AFAIK Atom in XOLO X900/Orange San Diego is a pretty good one. And a X86 option is always preferable to ARM in server enviroment.
"low-power high-density server systems packing as many as 2800 server blades into a single rack."
In other words - an ideal hardware for a hosting company to sell as dedicated server(s). Not really the "ultra highend".
I think I'd have to agree with MrJay. I'm not aware of the exact Atom model you're referring to and I'm sure it may perform better than the Calxeda but clearly for something with wattage that low, HP clearly didn't care about the extra performance of Atom. That assumes Atom has a better performance-per-watt, which I doubt.
The only reason why x86 is preferred in server environments is just because it's a little more compatible. People with serious servers tend to prefer PPC, SPARC, or even GPUs. ARM is perfectly capable as a cluster server platform, and linux is currently lightyears ahead (whether you want to look at that in either time or metaphorical distance) of pretty much any OS that is ARM compatible.
http://www.anandtech.com/show/5365/intels-medfield-atom-z2460-arrive-for-smartphones
That power consumption includes the graphics core, communications etc.
And now move up to Centerton, which is a dualcore Atom with 6W power consumption.... ARM is good for phones, but i don't see them as viable competitors for now.
That Atom is actually pretty impressive, seems like the first decent low-wattage CPU intel has made since the original atom (which was good for its time). I'm glad intel has finally taken low-wattage platforms seriously, because they always seemed to put them at lowest priority.
Although android is linux based, the changes done to android aren't always moved to linux, and the changes done to linux aren't always moved to android. I think its linux kernel 3.2 or maybe 3.3 where ARM has been getting from 25%-60%+ faster thanks to development from Canonical. I'm not terribly familiar with how Android works but last time I checked, it tends to use very outdated kernels, and if it hasn't received any of the 3.2 or 3.3 kernel updates, then that Atom CPU could actually be nearly head to head competing with those ARM chips. At that point the only difference would be licensing and price, in which ARM would likely win.
I noticed ARM seems to be a better competitor when it comes to devices that idle a lot and when you have a multi-core processor.
Plus HP didn't threw out the ARM solution, they just probably want to release the product ASAP, while the ARM solution is only on paper for now. According to what was written about these servers, later you can easily have ARM, Intel, AMD mixed according to your needs. The fact that Intel Atom is the first version talks more about availability than preference.
However, what I think won the argument for Intel simply was the readily available existing software that has already been tried, tested and proven on x86 platforms.
Wow, that means Android are using Old Kernel without the optimization from newest development? So the test from Intel Atom vs ARM wasn't really up to date?
Android 4.0 uses 3.0.8 kernel. Test was "up to date" of current Android status. By the way, on other side Atom was tested on 2.3, and that means even older kernel; it should get some boost by movin to Android 4.0 as well.
Moonshot, and other microserver/high-density server projects, aims squarely at cloud computing. Take Google Docs, for example. To make a page appear in Google Docs for a user to edit takes very little CPU resources; stick it on a Xeon and you're using an ICBM to crack a nut. Stick it on a microserver and you've got a much more efficient system - which means a system that is cheaper to run.
The trouble with migrating something like Google Docs to a low-power ARM or Atom server is in scaling: to make one page appear in Google Docs takes very few resources, but to make one million pages appear nigh-simultaneously is a different matter. The extra power of the Xeon chip helps, but it's still inefficient - especially as, when you get down to a low enough level, the processor can only execute a certain number of instructions at the same time. Let's assume a six-core Xeon with Hyperthreading - and further assume that Hyperthreading is actually suitable for this particular instruction, which it may not be: that's an absolute maximum of twelve simultaneous threads per chip. Four Xeons per 3U server, 13 servers per 42U rack (leaving 3U for switches) means 624 simultaneous threads and a whopping 4,680W TDP. That's 4.7KW you need to provide in power, and 4.7KW you need to cool.
Looking at Project Moonshot - and taking the Calxeda ARM chip's figure of 1.5W TDP for a dual-core chip - we're talking 2,800 'servers' per 42U rack, each with two processing cores, but no equivalent to Hyperthreading. Result: 4,200W TDP to power 5,600 simultaneous threads.
So, looking at cost per thread - and ignoring 'performance per watt' as we're making the (very big, and most likely fairly incorrect) assumption that loading a Google Docs page is simple enough that either chip can complete it in microseconds - you're talking 0.75W TDP per thread for a Project Moonshot server, compared to 7.5W TDP per thread for a traditional Xeon server. That means you can either support ten times the users (bandwidth allowing) for the same running costs, or reduce your costs to one-tenth while supporting the same number of users - or a combination, such as increasing your user base five-fold while still reducing costs.
You can all take those tinfoil hats of now, they really aren't watching anymore.
Intel are simply trying to dip their finger in as many pies as possible. I am sure HP have their reasons though.
Universal greeting to you to by the way (oops showing my age again)