the customer is always wrong
The customer tells you that they want cheap low power chips to basically glue the ethernet cable to the sata cable.
You tell them that yes - so long as the chips are 64bit, run Windows-TM and VMWARE.
As expected, Intel is closing out the year in the server arena with the launch of its "Centerton" Atom S Series processor, what the company's top brass is billing as the first Atom-based processor aimed at servers. The Atom S1200, as the first generation is known, is also the foundation from which Intel will build a bulwark …
I run an Atom D510 as a "micro" server. I found the problem to be the bandwidth bottleneck of the CPU/Motherboard combo.
For me, it peaks at about 180* Mb/s through any combination of ports (Ethernet; built-in SATA; RAID card on PCI-E; PCI card on PCI). Now while that may seem plentyful, certain combinations were even slower (PCI --> PCI-E).
While the specific number is probably a motherboard limitation, the limited cpu bandwidth (in the form of PCI-E lanes) and associated motherboard support is probably the problem. This issues does not appear to be addressed in this new generation cpus.
* - To those noting that the gigabit is only capable of about 120 Mb/s anyway, I'd like to point out that the maximum network throughput when reading from disk for me was just shy of 80 Mb/s, and the 180 Mb/s figure was from SATA to PCI-E RAID. Meanwhile, network card was perfectly capable to read to and from RAM at close to 100 Mb/s.
This post has been deleted by its author
Price comparisons are going to be interesting and possibly decisive for this. A year from now it might well be possible to get the equivalent of 100 ARMs (in all kinds of SoC and multichip designs) for the price of one of those Atoms and they will each be nearly as good at running Apache/nginx as the Atom and you get a free rain forest with the energy saved on the total system and you can rent the space you save in the data centre to the immigration department.
Well you need the 64-bit addressing so you can run VMware and only windows shops would think that might be a good idea.
What really gets me is the pricing on the a lot of small computing platforms. Atom SBC with video, ethernet and sata is being sold for $900 for an eval kit.
Come on HP, I want a micro-server on a PCIe card. 6 SATA ports and 2 Gig/E ports please - maybe a proper RAID option too.
Or you need 64-bit addressing to allow you to run large memory spaces and use these devices for caching.
Think nginx/memcached to scale web applications - CPU requirements are fairly low and if you want to maximise the performance, you have to avoid the relatively slow I/O buses on these CPU's.
Well, I want one of these for my always-on home server. (Which at present is an older Atom system that uses more watts for less performance).
With completely passive cooling and a solid-state disk drive, so zero noise. (current is passive cooled but not SSD)
And I get ECC RAM, so no data-corruption caused by faulty RAM
And I get to run VMware, so on the rare occasions I want Windows, I can have it without rebooting.
Unfortunately Intel's web-site doesn't yet reveal any S12x0 motherboard that I can buy, lt alone whether there's a mini-ITX one.
Intel still has the advantage of better process technologies than anyone else.
The interesting thing to watch will be whether Intel's process tech is sufficient to overcome all the historical baggage that an x86_64 chip has to carry around. The "green dream" would be ARM CPUs fabbed by Intel. It may come to that one day.
Also there's a large, though gradually diminishing, pool of software that isn't yet available to run on anything except the x86 architecture. That may be a lot more significant to business customers than it is to consumers.
Loads of server deployments don't need even 32 bit operations, never mind 64 bit. On addressing the overhead of extending beyond 32 bits is small. But again many server deployments don't need more than 4Gbytes and if you are running virtual servers on 16G RAM, then the overhead of any address extending scheme is small.
Supercomputers, Rendering, Modelling, Workstations etc can benefit from 64bits. But I just don't see the need if you are using multiple cheap server nodes each with own RAM.
"Intel knows that it's real customer is Microsoft - not the guy buying the chips"
Hello, 2002 calling, they want their argument back.
Here in 2012 server deployments are increasingly being rolled out without any Microsoft software at all.
You will recall that this chip is aimed at servers, right?
Even if intel succeed in getting these Atom chips into the data centre they still lose, just less than they would if it was an ARM cpu instead.
The bottom line is that for every Atom (or ARM) cpu that is put into a server there is a $400 Xeon that stays on the shelf.
As someone pointed out elsewhere, intel is not structured in such a way that they can survive on selling $50 chips.
Yes, the atom is 8 times cheaper. But on the other hand, the server need 10 times more CPUs to workout the same amount of workload (There is a comparison between Xeon and ATOM in the article). So, after all, Intel will earn more money if you choice the ATOM. There was a comparison between Xeon and claxeda ARM.
http://www.theregister.co.uk/2012/08/13/xeon_vs_calxeda_arm_apache_bench/
It can be see there by deploy the micro servers, company will actually spend more money on the server to cover the same workload.The benefits of microserver is that when the workload is I/O bonded or server is idle, the server saves power.
"the server need 10 times more CPUs to workout the same amount of workload"
Well yes, as you allude to, these are for servers that do not have high CPU workload requirements.
A lot of servers are not CPU bound and spend most of their time doing pretty much nothing.
In the past these would still run expensive Xeon CPU's because that was all you could get for servers, even if they just wasted cycles most of the time.
They could get away with this because they were the only game in town. You want a server running Windows and ECC just to serve a static webpage? Then by golly you must get this $400 CPU.
ARM and the fall of MS has changed that game.
Troll? OK, probably not.
A 33-bit processor would be very silly. It would mean that it had an Integer operations unit that could do 32-bit integer arithmetic and 33-bit pointer arithmetic. But fetching packed 5-byte fields would slow down the memory system compared to fetching aligned 8-byte fields. So the 33-bit integer operations would need to load and store 64-bit fields and fill the top 31 bits with zeros.
And then next year when the manufacturer wants to make a 16Gb version? All your 33-bit code would have to be recompiled and revalidated for 34-bit operation!
The only reason for 32-bit CPUs was that the cost of four extra RAM bytes per pointer was significant back in the days when we were migrating away from 16-bit CPUs. Today, the cost is not significant and 64 bits is the natural size for a pointer, that will be big enough for a LOOOONG time to come. (Hint: calculate the number of protons in the whole Earth. We won't need 128-bit addressing until we're well on the way to converting the whole solar system into computronium. )
Some people confuse the number of bits operated on by the integer processor with the size of a physical RAM address. That can be any number of bits. Virtual to physical mapping goes from a 64-bit number to an N-bit number. Most of the possible 64-bit virtual addresses cause access violations. That's also good: it makes it more likely that randomly corrupted or miscalculated pointers in a program get noticed and debugged! Cue remisiscences about the uses of 0xDEADBEEF or 0xDEADBABE or more boringly, IEEE Floating-point NaNs with the low-order bits set to the address they are stored in.
Someone doesn't know what BGA is.
It stands for Ball Grid Array. The balls in qustion are small solder spheres, by means of which the CPU is permanently melted onto its motherboard as per current Atom designs. In other words, the CPU in a Centeron server is not upgradeable (other than by replacing the entire motherboard). Since almost everything apart from the connectors and the RAM is on the chip, that's probably no great loss.
Nevertheless, a 6 watt chip that plugs into any i3 socket ,which doesn't need active cooling, would be welcomed by myself. (Yes, I do know if you spend £100+ on a box, you can find exotic cases that are in effect hundred-watt heatsinks, containing multiple heat-pipes to remove heat from a full-power desktop CPU)
(Well three comments -- first, good article).
Firstly, Intel won't have long to make headway with this Atom -- ARMv8 supports 64-bit, this spec was finalized in late 2011 and vendors announced silicon around October 2012 or so. Linux has gained 64-bit ARM support for v3.7 kernel (in the Linux tradition, I think it gained support before any silicon actually has shipped...) The info I read on these specifically said these'll run 32-bit apps under a 64-bit OS (if there's any closed-source ARM apps you can still run them; since Linux distros have supports mixed 32-bit and 64-bit on x86 I'm guessing it would not be a big deal on ARM) and supports virtualization. I don't know if it's out yet, but it should be out "Real Soon Now" if it's not already out.
Second, what is it with Intel and randomly adding and removing features? Why would they ship a 64-bit Atom without VT-X? My Atom Z520 *has* VT-X, so it's not like there's no Atoms with it. Weird.
>Second, what is it with Intel and randomly adding and removing features? Why would they ship a 64-bit Atom without VT-X? My Atom Z520 *has* VT-X, so it's not like there's no Atoms with it. Weird.
It's what's known as profiteering. A consequence of having unchallenged market dominance. Just as Nvidia invest in R&D of crippling parts of their GPU designs, then churn out the crippled derivatives at a low price to fend off competition while charging a hefty premium for the uncrippled originals.
In this case I think it's an arrogant miscalculation. Both AMD and ARM are capable of offering competitive products with a full set of features.