bit-tech.net

AMD targets Atom microservers with Opteron X-Series

AMD targets Atom microservers with Opteron X-Series

AMD's new Opteron X-Series, codenamed Kyoto, is an APU-based answer to Intel's microserver-targeted Atom S1200 family.

AMD has answered the immediate threat of Intel's server-centric Atom processors - and the somewhat more distant threat from ARM's multitudinous licensees - with a new Opteron family: the small-core Opteron X-Series.

Part of AMD's renewed focus on the microserver market - which began with its acquisition of SeaMicro an embarrassingly short period of time before its new subsidiary launched an Intel-based product - the Opteron X-Series is a range of low-power chips designed to pack as many cores into as small a thermal envelope as possible. As with Intel's Atom processors, and the small number of ARM-based microserver processors hitting the market, the emphasis with the X-Series is high levels of concurrency rather than raw performance.

Designed to compete with Intel's sever-centric Atom S1260, AMD's Opteron X2150 - codenamed Kyoto - has a 22W thermal design profile (TDP) and is the company's first accelerated processing unit (APU) system-on-chip design for the datacentre market. As a result, it packs 128 AMD Radeon HD 8000 stream processors alongside its four general-purpose processing cores each of which gets a dedicated floating-point unit and a 512KB chunk of the chip's shared 2MB L2 cache.

Compared to Intel's Atom S1260, the Opteron X2150 certainly has its advantages: as well as on-board GPU cores, which can be harnessed for general-purpose GPU (GPGPU) offload of highly parallel workloads, the chip supports up to 32GB of DDR3-1600 memory per socket - providing you can afford to splash out on 16GB DIMMs - compared to 8GB of DDR3-1333 for Intel's offering and includes four Jaguar-based CPU cores to the Atom's Centerton-based two. It also offers a claimed two-fold improvement in single-thread and overall throughput performance over the Atom S1260, when clocked to an equal 2GHz - although the chip itself will be launching at a maximum clock frequency of 1.9GHz.

As a system-on-chip design, the Opteron X2150 includes integrated support for eight lanes of PCI Express 2.0 connectivity, eight USB 2.0 ports, two USB 3.0 ports, two SATA III ports and DisplayPort, VGA and HDMI video outputs. Built on a ball-grid array (BGA) package, the chip measures 24.5mm by 24.5mm.

The X2150 is also to be joined by the Opteron X1150 CPU-only variant, which ditches the GPU portion of the design for workloads that don't require GPGPU offload. While this doesn't result in any reduction in package size - so as to keep the two chips interchangeable in original equipment manufacturers (OEMs) designs - it does allow for a boost from 1.9GHz to 2GHz and a reduction in TDP from 22W to just 17W. That peak TDP can, however, drop as low as 9W for the X1150 and 11W for the X2150, although AMD has not specified whether such a low power draw is possible under sustained workloads.

While the peak TDP of AMD's Opteron X-Series doesn't come close to touching the 8.5W peak of Intel's S1260, it does include twice as many integer processors - albeit with the same number of floating-point units - as the Atom chip, allowing for double the core density in the same footprint providing adequate cooling is available. The 17W peak TDP of the GPU-less X1150, adjusted for a per-core figure, is also an exact match for the S1260 - meaning AMD isn't as far off the mark as the raw figures might suggest.

The first company to use the new chips will, naturally, be AMD subsidiary SeaMicro, closely followed by HP integrating the processor into its Project Moonshot microserver product, but the company is clearly hoping the chips will have mass appeal in the burgeoning cloud computing market - and, if they perform as well as claimed, it could certainly have come up with something to wipe the smirk off rival Intel's corporate face, especially given the CPU-only X1150's similar pricing of $64 and the extra GPU grunt of the X2150 at $99 (both prices based on buying trays of 1,000 units.)

7 Comments

Discuss in the forums Reply
rayson 29th May 2013, 12:39 Quote
YAY AMD show them
Snips 29th May 2013, 15:30 Quote
"and, if they perform as well as claimed, it could certainly have come up with something to wipe the smirk off rival Intel's corporate face,"

Don't we get the exact same line with every AMD marketed release?

And what tends to happen again? oh yeah, it's all BS!
rollo 29th May 2013, 16:42 Quote
Cooperations no longer bielieve in AMD so no longer buy from them, Theres a reason they are below 10% on the server market now behind ARM even.

It will also take alot more than 1 good product launch to correct AMDs issues in this market.

Packing them with Onboard GPU is smart as in the low end they maybe able to get there chips inside home servers and save people a gpu purchase as well assuming its good enough to handle the general needs of that user.
LightningPete 30th May 2013, 01:32 Quote
If GPU processing can be used in more server market applications then this could be a winner in the microserver segment. However its their marketing department and actual test result/benchmarks using real world software that will see this go up or crash down
Goty 31st May 2013, 06:23 Quote
Quote:
While the peak TDP of AMD's Opteron X-Series doesn't come close to touching the 8.5W peak of Intel's S1260, it does include twice as many integer processors.

AMD's X-series processors also don't require an additional I/O chip, whereas the Atom chips do.
Chicken76 31st May 2013, 20:47 Quote
Quote:
Originally Posted by Article
...its four general-purpose processing cores which ... are split into two pairs each sharing a single floating-point unit and a chunk of the chip's 2MB L2 cache.
Where does it say that the FPU is shared a-la-bulldozer?
The L2 cache yes it is shared, but each core has it's own FPU I believe.
Gareth Halfacree 31st May 2013, 21:26 Quote
Quote:
Originally Posted by Chicken76
Where does it say that the FPU is shared a-la-bulldozer?
The L2 cache yes it is shared, but each core has it's own FPU I believe.
You're quite right - that was a misunderstanding on my part. I'll get that corrected, although as the CMS is still having a few hiccoughs it'll have to wait until the IT Elves have sorted things out I'm afraid - otherwise I risk making the whole darn article vanish!

EDIT: Phew: tested editing out on a never-published draft article, went fine, so I risked updating this article - and it's still there. Hoorah!
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