Slashdot Mirror
Samsung May Start Making ARM Server Chips
angry tapir writes "Samsung's recent licensing of 64-bit processor designs from ARM suggests that the chip maker may expand from smartphones and tablets into the server market, analysts believe. Samsung last week licensed ARM's first 64-bit Cortex-A57 and Cortex-A53 processors, a sign the chip maker is preparing the groundwork to develop 64-bit chips for low-power servers, analysts said. The faster 64-bit processors will appear in servers, high-end smartphones and tablets, and offer better performance-per-watt than ARM's current 32-bit processors, which haven't been able to expand beyond embedded and mobile devices. The first servers with 64-bit ARM processors are expected to become available in 2014."
I understand the implications of lower power for infrastructure reasons. Lower power means lower cost for power, lower cooling needs, etc. I get that. But what is the "Killer app" for these low power servers? Is it data warehousing? Simple web hosting? I can see these being useful for odds-and-ends servers in data centers with bigger iron for more heavy duty apps, but why is everyone jumping on this bandwagon?
Nobodies Prefect
Tidbits for Techs Technology Blog
Apple decided to stop buying their LCDs, why would they begin buying their ARM chips now?
Eh actually sorry, Samsung decided to stop selling to Apple, so why would they sell ARM chips now?
The new Google Nexus phones are shipping with 2GB of ram, and its conceivable that tablets will being shipping with > 4GB of ram within a few years. It just looks like Samsung is covering their bases for the future.
Bye!
..and sammy would probably license any chip arm makes anyhow, if only to just check it out.
world was created 5 seconds before this post as it is.
I would gladly welcome commodity server motherboards with ARM 64 bits CPUs. I trust I would easily find a suitable distro for my home server.
I hadn't the slightest objection to his spending his time planning massacres for the bourgeoisie... (P.G. Wodehouse)
There's no technical reason why an ARM chip of comparable performance to x86 could not be made. There's also no reason to believe such a beast would use significantly less power than an x86 chip, either. In order for the entire exercise to make any sense, they would have to target a niche between current ARM and x86. If they can keep the design sufficiently simple, they should at least be able to beat current ARM designs in performance and x86 in price. It is not clear that they can beat future Intel CPUs on power usage, especially since Intel's manufacturing process leads the industry by a significant margin.
While it is a workable hack to support more than 4GB of RAM without expanding the virtual address space, it is a hack. Much better to just g 64-bit and call it good. Hence I imagine that's what you'll see with mobile devices. When they start to need more RAM, they'll shift to 64-bit.
Same shit with desktops and many Intel servers. Intel supported PAE, and Windows implemented it as AWE, with 32-bit chips. Was never very popular though, due to the limitations and performance issues with paging. Now, with actual 64-bit chips, it has gotten much more popular.
I seriously doubt Apple will ever switch to ARM chips in OS X (not iOS) machines. They don't provide enough performance to run at the level of current OS X machines, not to mention that ARM64 is immature as hell.
No, but the threat of switching will provide that extra minute push to ensure Intel's continued refinement of Atom chips, and perhaps force them to release subsequent generations a year or two sooner than otherwise. Now that MS is actively promoting ARM-based tablets, Intel should be worried if not outright scared.
'He who has to break a thing to find out what it is, has left the path of wisdom.' -- Gandalf to Saruman
There is. The fact that ARM architechtule an order of magnitude or more behind the current x86 generation in terms of performance is a technical issue, and ARM is clearly having issues making its chips scale in speed without completely losing whatever advantage it has in low power. Hence all the talk about dark silicone in ARM world.
It is not clear that they can beat future Intel CPUs on power usage, especially since Intel's manufacturing process leads the industry by a significant margin.
Everybody says that, but it's only true for the high performance / high power consumption process variant. It's not true for the lower power variant(s), which have some differences and are more tricky than the high perf ones (I'm not an expert on this but one issue for example is that LP needs larger wires to reduce resistance and power consumption. This requires in turn more precision to avoid shorts between wires. People who know more on this topic, please share. It's important to understand how the race can turn in the low power area). For low power Atom chips Intel is right now on 32 nm, while TSMC has been on 28 nm for a while now. It's a one year and half-node advantage for TSMC clients. And Samsung is also now on 32 nm (par). Intel announced they will speed up the availability of new finer processes for low power in the future, but based on their respective announcements Intel and TSMC would be on par for LP (we'll have to see how this turns out in practice...). This means that ARM clients can have a competitive process in the low power space today, and possibly tomorrow. It's likely that ARM clients would focus on many cores / low power servers for I/O bounds loads. They can be competitive there, and gain a foothold. Going to higher single thread performance can come later, it would be hard to attack Intel there in the short / medium term anyway. If you pick a fight, pick one you can win. And the ARM world has more experience in LP.
Did you notice that the website has not been touched since 2010?
I am not sure how the design may progress with the current status. Is there any forks or active contributors?
An order of magnitude behind? No. A15 is close to Pentium M in terms of IPC. It should be around half way to Ivy Bridge IPC, I would think. That's not an order of magnitude, unless you're counting in base 2.
Indeed, with AMD's recent moves into the ARM ecosystem, I think Intel is pretty much the only one that isn't in on ARM. It remains to be seen what AMD's APUs are going to look like with an integrated ARM core, but I like the idea. I just hope they can pull it off.
It depends entirely on the application. For heavy maths processing in games or Photoshop ARM is way behind, but for typical server applications it is fairly competitive. Being low power is a huge advantage in datacentres and you often get better performance by having more cores than you do by having fewer faster cores.
Look at graphics cards. Lots of small, simple and not even terribly fast cores (in terms of clock speed). For that application they blow any CPU away. Now look at a typical server and you will see that it already has lots of small, simple and relatively slow cores dedicated to things like TCP/IP offloading, RAID array control and SSD management.
const int one = 65536; (Silvermoon, Texture.cs)
SJW, n: "Someone I don't like, and by the way I'm a fuckwit" - AC
MIPS and ARM are very similar Instruction Set Architectures. While I've only taking a cursory look at the new ARM64, it doesn't seem as clean as MIPS64. So with the same level of optimization, MIPS should be able to get a better performance per watt and higher IPC. SiByte had working MIPS64 CPUs 12 years ago. MIPS used to dominate the TOP500, but recently Intel has left them in the dust. So I don't see how Samsung is going to do any better with ARM.
Then again, if x86 can become the IPC leader, any ISA has a shot. Samsung is going to need a few Jim Keller's of their own to pull this feat off.
I just got an email from Samsung at the beginning of this week asking me to apply for a job in Dallas, TX. They are looking for an ARM Server hardware/software development team. They sent me the software job description and it looks like they want people who can tweak some firmware and perhaps even tinker with the Linux kernel. Looked like a great job but they require an MS in CS/CE and prefer PH Ds. I don't know why they even bothered emailing me. They have my resume and I clearly do not have a masters.
Kim Possible is the Supreme Kim. Cute as hell too.
Mod me up/Mod me down: I wont frown as I've no crown
Comment removed based on user account deletion
Given the fact that Microsoft will probably sell no more than a few thousand of those tablets (to all tech review sites) I'd say Intel is probably safe.
Your point stands though.
Write boring code, not shiny code!
Apple is busy switching to these chips for their laptops/desktops. Coincidence?
Yes. Because when I'm looking for a server, I need to be able to RUN THINGS ON IT! You know, like exchange and our autocad plugin licensing program and our fax controller and our voicemail control program and active directory. I can't even think of anything an arm chip would help with other than being slow and not running anything. It wouldn't even work as a file server because then it'd have to run some RT-style version of server 2012 and Linux probably won't run on it either. What a pointless device. It'd maybe work for a SAN or something but not a server.
... A couple of game servers, a Bittorrent tracker, IRC server...
You just compared a non-released chip from 2012 that is yet to be seen in any real life applications to a mobile x86 chip from 2003, and still didn't even get parity.
Half way to chip tech from 2012? Are you sure you want to lock that answer?
So what we get is that GPUs outclass ARM in simple mathematics, x86 outclasses ARM in more complex things.
So, where is ARM supposed to actually fit in that grand scheme of server chip? That is a question answer to which big companies are yet to find in spite of a lot of very vocal looking.
I look at the raspberry pi at $25 and think that would make an OK server. So a slightly upgraded pi with a good arm processor and say 4-8G of memory would be an awesome server as part of a cluster. For many servers in clusters a bit of storage is needed for boot up; the data mostly stays in ram. Then all that would be needed would be the occasional more traditional machine for HD storage. It would be killer to be able to keep adding new little servers for $99.
10 machines with say 4 cores each and 4G each would give a cluster with 40 cores and 35 gigs of in ram storage; all for around $1000. Plus anything by ARM would probably be pretty good efficiency-wise.
Due to redundancy and the extreme capacity adding flexibility I would much prefer $99 machines to just a boring regular server with just an big old ARM chip. Or even a boring regular server with a pile of ARM chips.
At hightly paralelizable symbolic processing. Like the GP said, that is most of what servers do nowadays: webserver, database manager (to a lesser extent), network filtering and caching, and a lot of other things.
I guess the bigest question is: How well will those new chips handle cryptography?
Rethinking email
Arm seems to be pretty strongly pushing linux support for 64-bit arm devices and I imagine the opensource server apps will pretty quickly follow. Samba4 will probablly run too if you want to host an AD tree.
It will be insteresting to see what MS does and whether they gimp the 64-bit arm version of windows in the same way they gimp the 32-bit one but even if they don't I wouldn't expect specialist propietry server apps to be ported any time soon. So for those uses arm is probablly out of the running for now.
note: i'm known as plugwash most places but i screwd up registering that here somehow in the past and now can't register
Game servers may be problematic because afaict they are usually propietary.
note: i'm known as plugwash most places but i screwd up registering that here somehow in the past and now can't register
GPUs mostly rule that roost (crypto, highly parallelizable simple tasks). ARM has nothing on them in that regard, it's about a decade behind just like it's about a decade behind in x86 in general computing.
Let's be serious here: ARM is something companies have explored for a while now. It just isn't up to par (yet?) If it were, we'd see a lot of applications as server market is very profitable. And considering both the GPU and x86 advancement, chances are ARM will never be able to catch up.
You make some great points here about process size, however one thing to remember is the Atom line processors are already 64 bit with very good performance to power draw ratios. Paired along side of Intel's graphics offers this allows for extremely low power / reasonably good performance devices to be produced in house.
I tend to agree with the others here who suggest this is a solution looking for a problem, that said, I'm always excited to see innovations which push Intel's glass prices down :)
That was my thought... it would be fairly easy to set up a cluster of arm CPU based servers as proxy servers, file servers, and even distributed database servers... Anywhere the main bottlenecks are IO, and you need to scale wide anyhow. I personally think the likes of non-relational database servers (NoSQL) as well as Node.js and similar asynchronous platforms cry out to be used by ARM systems. If your bottleneck is storage I/O throwing more CPU at the problem will not help. Right now this is roughly half of where the problem space lay.
That said, It would be nice to see a production ARM system platform that doesn't wind up being more expensive than the x64 equivalent. Virtualized servers on x64 each connected to distinct drives by the host currently work out better in terms of cost than the wide scale systems that the current ARM server systems are starting to offer. Raspberry Pi shows that it can be done economically, but the biggest limitation there is simply the amount of memory... if you could get something similar to an R-Pi with 2-4GB of memory, and even a single SATA port, you could create clusters of systems that could scale wide very economically.
Michael J. Ryan - tracker1.info
But you can run a number of smaller virtual environments in the larger x86 (64-bit) CPUs that wind up scaling very well in comparison to a small cluster of ARM systems... When I've priced out the ARM server systems currently available, x86 usually winds up being a better value, when you consider having a host + 1VM per compute unit, which still out-paces the ARM systems. That said, I think that some standardization of ARM based general computing systems could bring significant reduction in pricing which could lead in ARMs favor. I think if AMD and Samsung came together on a BIOS/UEFI common interface system for components, that general purpose ARM computing can really take off.
Michael J. Ryan - tracker1.info
I don't think MIPS per CPU is the biggest concern in datacenters, it's MIPS per Watt. In that sense, ARM may outclass x86.
"It's too bad that stupidity isn't painful." - Anton LaVey
You don't need a lot of performance in a laptop. Desktop sales have plummeted and Apple has negligible server sales. Make of that what you will.
That's the problem. At the moment, it's not outclassing x86 nor GPUs due to severely lacking in terms of raw calculating power. The word is that "there's potential if ARM can increase its performance while keeping the low power requirements." That doesn't mean "it's happening" or even "it's going to happen".
And intel isn't exactly sleeping, so while ARM needs to get to increase it's workload per time, intel is working on increasing workload per energy. ARM may potentially hit the "sweet spot" at some time before intel as they are both targeting the same thing but from different angles. Or it may not, as intel gets there first. Either way, neither of them are there yet, in spite of hilarious hyperbole often pushed in the media looking for sensationalist headlines.
By all means, what do you understand by "symbolic processing"?
No, GPUs don't rule at that. (And crypto is not symbolic processin, thus it is an open question.)
Rethinking email
Eh actually sorry, Samsung denied reports it decided to stop selling LCDs to Apple.
https://en.wikipedia.org/wiki/Symbolic_computation
Why is ARM better then x86 or GPU for this particular task, and for that matter, why is it so important and cost intensive that you would look to have gains after software rewrite to ARM? I honestly have no clue, enlighten me.
Were you the guy who explained why the conditional instructions of 32-bit ARM were no longer an advantage, hence the lack of them in 64-bit ARM? I forgot the explanation and can't find it, and would be grateful if you point me to it.
Stick Men
How to spot an article written by a matematician: That definition is perfectly correct. Altough, a lay-person will be misled by it, thinking that symbolic computation is just one of its applications (the one the author uses more).
The same way that a computer can only work with numbers, but we abstract those and think about an image or sound wave, we also often abstract the symbolic algebra, and think about text and data objects.
GPUs are built for numerical computation. They are just terrible at branching, and thus, can't deal with complex data.
ARM has smaller cores than x86, that use less power each. That means you can fit more ARM cores at the same space and same price. If the performance is comparable (what we don't know yet, but isn't an unreasonable assumption), an ARM chip will be able to process hightly paralelizable symbolic tasks faster than an equivalent priced x86 chip.
Rethinking email
I'm not planning to install a datacenter in the near future, so I haven't studied it in-depth. But I'm sure the big players will do just that, and if they do make the switch, it won't be for fanboy reasons, but for hard economic facts. So I'll let that be the proof, and postpone any poor judgement from my part until later notice.
The only thing I can say is that from my experience with ARM programming (including assembly), the architecture is much cleaner than x86. And a clean design has a tendency to be more efficient. Not much to go on, I agree, but it indicates to me that ARM might have a lot of potential.
"It's too bad that stupidity isn't painful." - Anton LaVey
Ah, you're talking about hypothetical applications. ARM seems to be the king of those.
Shame no one ever managed to get those to work at levels needed for actual production so far.
The issue isn't about how "clean" architecture is, but how hard it will be to rewrite and/or recompile all your software for ARM. That is, assuming there are actual advantages to ARM.
So far, no one seems to have bothered in spite of a lot of noise about the issue. This strongly suggests that this is in fact just noise, and ARM is either not yet ready, or is generally unsuitable for server applications.
I was just addressing the issue of power consumption. I assume that an architecture that's basically horrible in terms of design will have some implementation and power issues (as a rule of thumb).
The recompilation issue is entirely different, but an interesting point to bring up. Is it really that hard? From my experience, it isn't. The major difficulty in cross-compiling is with different environments, not the CPU architecture (except for the very lowest levels). But if you're dealing with server farms, e.g. linux servers, the work has already been done, linux runs fine on ARM.
The only possible hurdle I can imagine is with optimized code that optimizes for cache hits on a specific processor. But these kind of optimizations tend to break even between different versions of the same architecture. I don't know how widely they are used in datacenters.
"It's too bad that stupidity isn't painful." - Anton LaVey
you can fit more ARM cores at the same space and same price. If the performance is comparable (what we don't know yet, but isn't an unreasonable assumption)
Isn't it unreasonable? Let's see.
Let us assume that an ARM core occupying less space, consuming less power, and cheaper has comparable performance to an x86 core. This would mean that large server farms (facebook, Google, Amazon and others) have an advantage by switching over to ARM. None have switched even though they are very concerned with space, power and cost.
Hence our original assumption was wrong. Which means that an ARM core does not have comparable performance to an x86 core.
QED
Bingo Dictionary - Pragmatist, n. A myopic idealist.
Samsung didn't start to produce this chip yet, how do you expect Facebook, Google, Amazon and others to use it?
Rethinking email
The Nexus 10 uses a Samsung made A15 based processor. A15s are showing up in consumer hardware.
So your comment was specific to Samsung processors? Do you think Samsung will do something no ARM processor manufacturer has been able to do? By a significant margin? Why?
Bingo Dictionary - Pragmatist, n. A myopic idealist.
Did you notice that TFA is about a new architecture from ARM?
The comment was not only about Samsung, there are other manufacturers tooling their fabs for this new architecture. But nobody is producing it yet.
Rethinking email
Yes but the new architecture is only an evolutionary one. There is nothing to suggest it will suddenly make ARM "comparable" to x86 in any sense. If new architecture ARM is comparable to x86, old one is as well because the new in in turn comparable with the old.
Bingo Dictionary - Pragmatist, n. A myopic idealist.
Ah, interesting, I saw the other article mentioned a lot around, but no one seemed to want the rest of us to know they denied it.... I'd mod you Insightful if Slashdot wasn't too paranoid about bias to let me do so.