ARM And The Server: A Sooner-Or-Later Inevitable Future, Or Simply A Deluded Fervor?
Here’s some irony for you; I’m sitting here writing this post in the midst of a technical session on Intel’s next-generation Sandy Bridge x86 microarchitecture. More from me on Sandy Bridge soon, I promise; right now, my focus is on ARM microprocessors. As any of you with business savvy already know (and I know that there are plenty of you out there), two fundamental means exist for a company to expand its business:
- Attract new customers who aren’t currently buying products in a particular business segment, or
- Steal away customers from your competitors in that business segment
This week, as was the case at the past few Intel Developer Forms, I’m hearing in abundance about Intel’s ongoing attempts with the Atom CPU to establish a beachhead versus ARM in handheld devices. Conversely, ARM is attacking traditional Intel computing strongholds; in netbooks with so-called ’smartbooks’, and more recently at the other end of the processing-horsepower spectrum with servers.
To date, ARM’s been ‘all talk and no walk’ in servers, in no small part because its partners’ currently available products aren’t really appropriate for this particular market. By means of review, subsequent to the unveiling of the ARM11 architecture, the company chose with follow-on Cortex to split the product line into three separate application vectors:
- Cortex-A: highly integrated application processors, primarily for mobile devices
- Cortex-M: cost-sensitive processors for traditional microcontroller applications, and
- Cortex-R: high-performance processors for deeply embedded realtime applications
Focusing on Cortex-A (which today constitutes the bulk of Cortex licensee shipments thanks to Apple and other high volume implementers), the existing family consists of the Cortex-A5, Cortex-A8 and Cortex-A9. Cortex-A5 is an in-order, single-issue architecture spanning 1- to 4-core implementations, whereas Cortex-A8 (exemplified by, for example, Apple’s likely Intrinsity-designed and Samsung-fabricated A4 processor) is also in-order (although a typo on ARM’s website suggests otherwise). Cortex-A8 at least to date also comes only in a single-core implementation, although its superscalar nature allows for dual-issue instruction execution capabilities in some code sequence situations.
Cortex-A9, as Ars Technica’s Jon Stokes noted in a recent writeup, further ups the feature set ante by not only allowing for multi-core configurations but also augmenting each core with out-of-order execution abilities. Its low-power attributes are certainly attractive in the heat-adverse enterprise, but it still lacks key features generally deemed necessary in this particular application:
- >32-bit addressing support
- Hardware hooks for virtualization acceleration
- Parity (or even better, ECC) support for both integrated cache and system memory
- Double-precision floating point arithmetic logic, and
- Multi-CPU cluster configurations
Cortex-A15, which received its first public mention more than a month ago via its ‘Eagle’ project name when Texas Instruments announced itself as the premier licensee, and which had its official coming out party last week, addresses the bulk of these current-generation setbacks (and potentially all of them…ECC system memory support is not yet documented but may exist). Targeting 32 (and below) nm processes and therefore not likely to ship before 2012, it’s forecasted to come in up-to-four per-die core counts and achieve clock speeds up to 2.5 GHz.
Will Cortex-A15 find use in traditional ARM markets, such as premium smartphones and PDAs, tablets, handheld game consoles, and the like? Sure. But judging from the feature set, ARM clearly (at least to my eyes) also has its sights set on Intel’s (Xeon) and AMD’s (Opteron) server turf. Will ARM succeed? On the one hand, I’m admittedly skeptical; consider that one month back I was negatively critiquing Micron’s latest enterprise-targeted SSDs because they didn’t support the SAS command set. Now consider how much more herculean a task it is to rewrite an entire operating system and program suite to support a new microprocessor architecture, versus simply converting an enterprise application or few from the SCSI to ATA command sets.
On the other hand, I can’t disregard that, for example, Apple essentially recompiled (and in the process rewrote large portions of) OS X first to migrate it from PowerPC to x86, then subsequently to develop the ARM variant for the iPhone and iPod touch which later became known as iOS. Enterprise power consumption concerns are indisputable and increasing with time; if ARM can make a substantial enough difference versus x86 in this regard, it’ll justify the necessary time, cost and effort investment to do the code port. At least one server company, Austin, TX-based Smooth-Stone, has taken the data center-on-ARM plunge. Then there’s Google, who in April acquired secretive startup AgniLux (formed from employees who left P.A Semiconductor after Apple bought that company in 2008, therefore with a mix of ARM and PowerPC in-house expertise). And job listings suggest that Microsoft may also be be evaluating server-on-ARM opportunities.
Consider, too, that with its SM10000, SeaMicro has constructed a datacenter server out of up to 512 Intel Atom processors. Granted, they’re x86 instruction set-based like their Xeon bigger siblings, but they dispense with extended-address support, high-performance floating point capabilities, ECC cache, and the like. If there’s sufficient power consumption motivation to architecturally back-step to this substantial degree, is it really such a stretch to imagine a further transition to a competely different instruction set? What do you think, readers?
steve commented:
If ARM aimed at running the hypervisor not the OS, it may be simpler to make a change from x86. The hypervisor makes adjustments to the unending variations in OS drivers and presents a more standardized interface to the CPU. Looks like they already put the hooks in for virtualization.
Charles Q. Tran commented:
Competition is healthy and good for all of us. ARM and its worldwide partners are ready for first round, then second round in 2011 and Beyond. Welcome on board.
Sofianitz commented:
MIPS has a better, more register-oriented RISC architecture than ARM. Especially for “lots of process switching” server computation. Pity.
logiotek commented:
Everyone who commented here is correct but I agree the most with what “technotart” has said.
Baskaran Kasimani commented:
Intel processors are just overkill in terms of power to perform even simpler tasks like serving the files over the network. It is something similar to running a big power plant to support a automatically run toilet.
May be someone has to look at the basics and go with a FPGA based design with low power targets to perform mundane tasks like serving files.
Charles Q. Tran commented:
Upfront Disclosure: I do not work for ARM and have been successfully implement ARM based products during the past 10 years.
Competition is healthy and it is perfect time for ARM to bring innovation to good old market.
Ian commented:
Up front disclosure: I work for ARM.
Nice article. 100% agree that it is perfectly reasonable for people to be skeptical of ARM's play here. As I told the audience at June's Structure event in San Francisco, ARM has zero market share in this space today. It will remain at that level for a while. It is a long term play that ARM and several silicon partners are working on. This is a complex problem beyond simply having the appropriate processor cores available. Significant software elements (although fewer than you might initially imagine) exist so customers will need to see a significant business value prop to cause someone to switch from incumbent architectures/platforms. That is what myself and my team are working on day and night.
technotart commented:
POWER CONSUMPTION
-----------------
Believing that using ARM will inherently save power is a delusion!
ARM processors are lower power than Intel processors because they are slow (at least if compared on the same process node). If all of the circuit tricks and speed optimizations that are used by Intel to get to high performance were used on an ARM implementation, you would be looking at a very similar power number.
Saving power in servers is about system architecture - not processor architecture. It is probably true that current Intel based server architectures have never really focussed on this by looking at re-architecting the whole system. Coming at the problem fresh with either processor architecture could yield massive power reductions through the judicious use of hardware acceleration for common data manipulation tasks. - this always being a trade-off between decreasing power and increasing software complexity.
Just Curious commented:
Doesn’t Moore’s Law dictate that ARM cores double in performance every two years? Seems inevitable to me that ARM cores will progress up the performance curve into servers. Intel’s challenge is that they are fighting against Moore’s Law on the low end, and have not yet found a high-volume profitable business model to replicate their previously very successful PC strategy on the high end.
ARMwrestle commented:
As Chris commented, most servers run Linux which has been heavily ported to the ARM architecture already and is in effect ready to go. The other aspect to consider is that the server space is not monolithic - there are many types and applications for servers from high-end data centres to emerging low-end SOHO & home servers. ARM does not have to win them all in one shot, but it can create a bridge head with its silicon partners and prove itself in market.
plebeian commented:
All I can say is I’ve been derided over the years for opining that ARM had a very significant future - the x86 crowd just complained it would never do ‘xxxx’. Now how many mobiles have x86 architecture in them? My own work environment is split between those who can’t see beyond Wintel and those of us (my bit) that has moved to ARM and Linux for our new product set (that the customers seem to appreciate!). Moving to ARM servers may be a challenge but it will happen - our modest server room uses near 10kW to power the hardware and another 5kW to cool it. What an irresponsible waste that is!
BobsUrUncle commented:
Who wants to be the first CIO to recommend a less than capable server infrastructure to their organization?
BobsUrUncle commented:
Let's list the number of failed chip companies that litter the server space:
MIPS
PA Semi
Dec
Sun
HP
now take a company with almost no experience in the server space. sell you ARM stock now while u have a chance.
Chuck commented:
Since at least 99.99% of the code in an OS is written in C++, the instruction set does not matter. One just simply recompiles the source for the appropriate CPU. When it comes to competing with Intel on cost, then most chip suppliers are at a huge disadvantage. If the "power card" gives a competitor an advantage in the market place, I am sure Intel can develop a lower power server chip that may or may not have lower performance. The clock rate is about maxed out. The real problem is accessing memory when cache is missed. Finding the sweet spot for power, cache size, instruction clock rate, and a better DRAM memory architecture is the real name of the game.
Not Likely commented:
If it were so easy to displace the x86 ISA then MIPS and many others would still be around. The question also is how big the non-x86 restricted market is. i.e. The roll your own code, java, scripting etc. How much of the cell phone market can be leveraged into a server solution?
Chris commented:
Also, remember that many servers and data centers use Linux which is already ported to the ARM architecture, so not such a big software task for some.
technotart commented:
Processor architecture almost completely irrelevant in this discussion. It is all about business models:
ARM’s advantage: It licenses cores at relatively low cost and enables any semiconductor vendor to compete with Intel in any chosen market segment.
Intel’s advantage: Process technology and manufacturing expertise. Nobody can match the sheer performance that Intel can deliver. If the end market can tolerate less than ‘bleeding edge’ single core performance through the use of other system architectures, the Intel can loose.
