HP to do ARM based servers
Major blow to Intel with Lenovo doing MIPS servers…
Oct 26, 2011 in Chips, Microprocessors, Servers
According to a report from Bloomberg News Service HP (NYSE:HPQ) will start manufacturing servers based on the ARM architecture in a sharp departure from its previous Intel-only design philosophy.
The processors for the HP servers will come from the startup Caxeda, which is partly owned by ARM. Caxeda is planning a quadcore processor based on the ARM Cortex-A9 design.
HP is not the first company to do ARM based servers. More than a year ago we reported that Facebook is doing the same thing, but the main difference is that the servers from Facebook are for internal consumption only.
This comes at the same time as sources have confirmed to SemiAccurate that Lenovo is mulling servers based on a MIPS design. The Lenovo servers that would, at least initially, be marketed in China would be based on a locally designed and fabbed chip also known as Godson.
Lenovo is mum on this and several requests for comments have not been answered by Lenovo. The departure from the x86 only architecture by both HP and Lenovo is a major blow for Intel, which has been the only supplier, together with AMD, to deliver chips to the two companies.
However, a major effort still remains to port software that is written for the x86 architecture to have it run on processors based on the design from ARM.S|A
The company is Calxeda (note the “l”)
This will be a mean angry bird server.
The interesting question is “why.” ARMs are only really interesting in servers on a performance / watt basis, but a lot of that territory has already been claimed through virtualization. It was mentioned above that porting Linux software and apps should be straighforward and that’s true enough, although optimizing them may be a bit more interesting than implied.
The only attraction I see for ARM servers are for relatively small, not-quite-embedded-level tasks (souped up NAS box, etc).
I thought Facebook denied that they were making ARM servers.
they just hate Intel for ditching Itanium processors
However, a major effort still remains to port software that is written for the x86 architecture to have it run on processors based on the design from ARM
Not necessarily. On these “simple” servers you do not run particular proprietary SW, but mostly services that are already available as part of OS or as components for it. I think about web servers and PHP which just need an optimized Linux kernel and no porting of other code at all. As well as printer or filesystem/backup servers.
BTW, most of the projects today are written in C/C++ or high level languages, so the effort of the porting is just the one required for pressing the recompile button.
[ SOS ] Complaint about Human Rights Violations by IBM China on Centennial
Please Google:
IBM detained mother of ex-employee on the day of centennial
or
How Much IBM Can Get Away with is the Responsibility of the Media
or
Tragedy of Labor Rights Repression in IBM China
In other news HP will be manufacturing servers that are really slow, highly proprietary, and require a convenient yearly service contract. That’s ARM for you.
“Intel, which has been the only supplier, together with AMD, to deliver chips to the two companies”
You cannot be only one together with someone else. It makes you one of two.
It’s even more of a blow to AMD.
how can it be a blow to AMD.
they dont have _any_ server marketshare to lose.
It’s not bad news for AMD because the recent Godson 3A-based computers have used Hypertransport and AMD chipsets.
So I’d think mixing AMD and MIPS could create a brand of alchemy not seen in years.
> So I’d think mixing AMD and MIPS could
> create a brand of alchemy not seen in
> years.
Ugh! What did we do to deserve to be pun-ished. :-)
(For those who do not know Alchemy was a
[MIPS ISA] processor brand at one time produced by AMD.)
Yeah, because the entire world gave AMD such a hard time for only being 80%-110% as good as Intel’s latest.
ARM chips with 1/16 the performance are going to knock AMD down another notch. Apparently AMD’s real crime is being AMD.
Unfortunately ARM chips do not have 1/16 of the performance of the fastest of AMD chip, but they are 10 times as efficient.
In servers where you have low computational peaks, you just compensate the lack of total raw power with more chips/cores. If they cost less and consumes less when you have reached your raw power threshold, you have the best solution.
ARM is quite able to offer it. Would it have 64-bit support, many AMD/Intel servers would soon be replaced. This will hopefully happen in a couple of year when an ARM with 64-bit capacity will be put on the market.
“10 times as efficient”
Source.
HP’s servers aren’t exclusively x86; they still have Itanium lol.
Let’s hope that HP gets their hands on 64 bit ARM CPUs soon. A server with only 4 GB RAM is not very exciting, LPAE or not.
For lots of servers 4 GiB of RAM is more than plenty. Not every server is a monster monolithic Oracle database or such…
Even smartphones have 1 GB RAM nowadays, concept phones like the 2 GB HTC Eerie HD3 showing the way. Unless HP is aiming for the embedded tiny server market, the lack of address space is going to hurt them.
Lenovo is in a better situation, the Loongson is a MIPS64 CPU.
Oracle? This is that Oracle working with
http://www.javaoracleblog.com/java/16_Processor_ARM_system.jsf
64GB 16-processor ARM device
Let’s be very clear. 4gb or RAM is not a limitation of 32 bit, it’s an artificial limitation of Microsoft 32 bit consumer operating systems.
Any 32 bit Linux will work with 64GB of RAM if you have it.
32 bit Windows Server 2000 Datacenter Edition supports up to 128GB or RAM.
MS could have easily added the real PAE code from WinServ2k DCE to 32bit XP, Vista, 7, etc… but they saw it as an opportunity to bilk you out of more money. Don’t assume that Linux on ARM can’t do what MS meerly chose not to do.
Au contraire mon ami; the 4GB address space barrier _ís_ in fact a limitation of 32bits arcs. PAE is only a workaround; it adds 4 bytes more to the physical address field and another page table level, enabling 64GB of RAM for the system. But in order to use it the OS has to make heavy use of double buffers and mmap() or similar concepts like AWE, slowing computations down and causing driver instability; especially audio and video drivers weren’t coded with addresses above 4GB in mind.
Besides, x86-64 has been around since 2003 and XP64 since 2005. At that time only the server segment chipsets supported more than 4GB RAM; desktop segment chipsets ranged from 1 to 4 GB RAM. Nowadays everyone with a half a brain is getting himself a 64bits hardware and OSses, so one shouldn’t feel sorry for them luddites.
typo: that should read “add 4 _bits_ more to the physical address field”
I never said it didn’t incur a performance hit, nor did I ask you to try to explain how it works.
However, ARM CPUs are hardly memory limited by DDR3 anyways, so any theoretical loss is unlikely to hurt a real-world workload.
PAE doesn’t remove the 4GB address space limit though, so processes are still limited to a 4GB address space, or relying on some form of paging. PAE just allows more physical memory to be addressed.
As I understand it PAE isn’t supported in 32bit consumer windows for historic reasons, design decisions made prior to the PPro being available mean most drivers won’t work properly.
“However, a major effort still remains to port software that is written for the x86 architecture to have it run on processors based on the design from ARM”
I’d be suprised if that was an issue for the kind of servers HP will be selling. Linux and the BSDs are well supported on ARM, and most software isn’t written down to the assembler level, so a simple compile should suffice.
That’s true.
But I don’t believe that the kind of workload that someone would run on an ARM based server is going to need 20GB of memory consumed by a single process.
I would assume that someone like Google would run Apache servers on ARM, trying to increase the ratio of cores to requests per second by using smaller, simpler cores, rather than having a large, complex core trying to serve many requests per second. You could probably fit about 40 ARM cores (sans SOC peripherals) in the same space as a single Xeon core, it might just make sense.
then there’s marvell’s upcoming aramada chip that charlie had previously written about, which is probably more impressive.
g vAMD’s 28nm Southern Islands-based Radeon HD 7000 GPUs have already entered production
Production as in consumer market?