bit-tech.net

AppliedMicro launches first 64-bit ARM server chip

AppliedMicro launches first 64-bit ARM server chip

AppliedMicro has become the first company with a true 64-bit ARM implementation aimed specifically at the server market.

Applied Micro Circuits has become the first company to announce a fully-functional server platform running on a 64-bit ARMv8-based system-on-chip processor.

While ARM is causing a stir in the microserver market with its low-power chip designs, in particular the Cortex-A15 'Eagle' which includes 48-bit memory addressing and hardware virtualisation extensions, the company still has a major roadblock to its major adoption: unlike x86 chips, the ARM processors are 32-bit at heart.

For ARM's traditional target market of embedded and mobile computing systems, that's not a problem: no phones on the market today have more than 4GB of RAM, and they don't spend their time doing heavy integer mathematics. If the company's serious about breaking into rival Intel's stronghold market of the datacentre, however, it's going to have to address that problem with a true 64-bit implementation - not just the extended addressing of the Cortex-A15.

ARM has announced that it's working on 64-bit chips, but actual silicon has been thin on the ground. That is, until now: ARM licensee AppliedMicro has announced a 'server-on-a-chip' solution dubbed X-Gene, which it claims is the first 64-bit ARM implementation designed specifically for the server market.

Originally announced back in October last year, X-Gene chips combine multiple ARMv8 symmetric processing cores featuring L1, L2 and L3 caches with a high-performance memory controller, an integrated Ethernet network interface and other communications interfaces to make the chip suitable for immediate implementation.

The best hardware in the world is no use without software, of course, but AppliedMicro claims to have that sewn up too with the first LAMP - Linux, Apache, MySQL and PHP - stack to run on a 64-bit ARM core.

'This is the first time the world is seeing a mature, fully-functional server platform running a real-world application on 64-bit ARM-based processor,' boasted Vinay Ravuri, AppliedMicro's vice president of processor products, at the announcement. 'As a result, AppliedMicro has already secured key strategic customers and partners around the world and has been enabling them with the tools they need to get started in advance of silicon. This web server emulates a live content delivery application featuring rich video, audio and text, and demonstrates the robustness and readiness of our next generation cloud server solution.'

So-called 'cloud servers' - which focus more on running massive quantities of relatively lightweight processes, compared to traditional servers' focus on rapid computation of a smaller number of more complex processes - are considered by many in the industry to be the next big growth area. It's something which has AMD concerned enough to splash out $334 million buying microserver specialist SeaMicro, while Intel is working to convince OEMs to use its low-power Atom chips for such devices.

With ARM's customers now shipping true 64-bit implementations, the British chip design giant is likely to rapidly increase its presence in the market - and that's something which will likely have x86 giants like AMD, Intel and VIA concerned.

13 Comments

Discuss in the forums Reply
iwod 27th April 2012, 11:57 Quote
So is it AdvancedMicro or AppliedMicro ?
Gareth Halfacree 27th April 2012, 12:05 Quote
Quote:
Originally Posted by iwod
So is it AdvancedMicro or AppliedMicro ?
Bah! Applied, not Advanced. My fault - I was thinking of the AMD (which *is* Advanced Micro) story I was writing afterwards... Fixed now!
r3loaded 27th April 2012, 12:21 Quote
Is this a custom implementation of the ARMv8 architecture, or based on an ARM reference design?

Sent from my GT-I9000 using Tapatalk 2
Gareth Halfacree 27th April 2012, 15:02 Quote
Quote:
Originally Posted by r3loaded
Is this a custom implementation of the ARMv8 architecture, or based on an ARM reference design?
Every ARM chip in existence is based on an ARM reference design - it's just a question of whether it's hard macro or soft macro. In this case, it's soft macro - so yes, it's about as custom as an ARM chip gets.
schmidtbag 27th April 2012, 15:38 Quote
unless there's a significant performance difference when running 64 bit ARM, i think i'd rather just use PAE. PAE in linux doesn't really have any noteworthy performance differences, but you're not limited to 4gb.
Gareth Halfacree 27th April 2012, 15:40 Quote
Quote:
Originally Posted by schmidtbag
unless there's a significant performance difference when running 64 bit ARM, i think i'd rather just use PAE. PAE in linux doesn't really have any noteworthy performance differences, but you're not limited to 4gb.
Existing ARM server chips - by which I mean the Cortex-A15 - don't need PAE, 'cos they use 48-bit memory addressing (despite still being 32-bit chips.) Each chip supports up to 1TB.
KarlFreund 27th April 2012, 17:03 Quote
Let's be clear here: this is an announcement of a working FPGA, not a working chip. But is great to see that the 64-bit ecosystem can now begin the work to get ready for real silicon when it DOES become available. Make no mistake: ARM is coming to the datacenter, and there are indeed plenty of 32 bit applications and workloads (ever hear of Java?) that will start the move 1st (with help from Calxeda!)

-Karl
Combatus 27th April 2012, 23:04 Quote
Gareth, I have to admit I just love the way you make hardware virtualisation extensions and heavy integer mathematics sound refreshing and interesting!
dicobalt 27th April 2012, 23:38 Quote
ARM on a server seems like throwing rocks at a tank. If your hardware and processing demands are that low you would be better served by virtualizing it on an existing machine and save the cost of expensive proprietary ARM servers.
Gareth Halfacree 28th April 2012, 00:10 Quote
Quote:
Originally Posted by Combatus
Gareth, I have to admit I just love the way you make hardware virtualisation extensions and heavy integer mathematics sound refreshing and interesting!
I'm going to choose to assume you meant that seriously. :p
Quote:
Originally Posted by dicobalt
ARM on a server seems like throwing rocks at a tank. If your hardware and processing demands are that low you would be better served by virtualizing it on an existing machine and save the cost of expensive proprietary ARM servers.
You're missing the point: the servers these are aimed at do lots of small tasks at the same time; in the power envelope of 16 x86 cores, you can get 512 ARM cores - and thus run 498 more threads. (Numbers pulled from you-know-where for illustration.) Doesn't matter how much you virtualise, if you've got sixteen cores then you're only running sixteen threads simultaneously (modulo tricks like Hyper Threading.)
r3loaded 28th April 2012, 10:47 Quote
Quote:
Originally Posted by Gareth Halfacree
I'm going to choose to assume you meant that seriously. :p You're missing the point: the servers these are aimed at do lots of small tasks at the same time; in the power envelope of 16 x86 cores, you can get 512 ARM cores - and thus run 498 more threads. (Numbers pulled from you-know-where for illustration.) Doesn't matter how much you virtualise, if you've got sixteen cores then you're only running sixteen threads simultaneously (modulo tricks like Hyper Threading.)
Yes, and an ARM server like this with (relatively) low performance per core but a high core count makes it especially suitable for tasks that are easy to parallelize but simple to compute such as a web server. Each individual request is straightforward to handle, but handling 500 of them at a time is a lot easier on 500 cores than time-slicing them across 16 cores.

For computationally intensive work, Xeons obviously still remain the best choice.
ch424 28th April 2012, 11:51 Quote
Quote:
Originally Posted by Gareth Halfacree
Every ARM chip in existence is based on an ARM reference design - it's just a question of whether it's hard macro or soft macro. In this case, it's soft macro - so yes, it's about as custom as an ARM chip gets.

No, there's three ways of licensing: hard macro (the customer gets a silicon layout that they can send to a foundry), RTL (the customer gets source code they can plug into their own cell libraries and turn into a hard implementation themselves, what you call "soft macro") and architecture license (eg Qualcomm make their own CPUs from the ground up, just using the ARM instruction sets). So to answer r3loaded's question, yes, this is probably a custom implementation, not derived from a reference design.
Quote:
Originally Posted by Gareth Halfacree
Existing ARM server chips - by which I mean the Cortex-A15 - don't need PAE, 'cos they use 48-bit memory addressing (despite still being 32-bit chips.) Each chip supports up to 1TB.

LPAE is 48-bit physical addressing.
Gareth Halfacree 28th April 2012, 19:38 Quote
Quote:
Originally Posted by ch424
No, there's three ways of licensing: hard macro (the customer gets a silicon layout that they can send to a foundry), RTL (the customer gets source code they can plug into their own cell libraries and turn into a hard implementation themselves, what you call "soft macro") and architecture license (eg Qualcomm make their own CPUs from the ground up, just using the ARM instruction sets). So to answer r3loaded's question, yes, this is probably a custom implementation, not derived from a reference design..
You're quite right - for some reason I completely forgot about architecture licensing.
Log in

You are not logged in, please login with your forum account below. If you don't already have an account please register to start contributing.



Discuss in the forums