AppliedMicro Announces 64-bit ARM Based X-Gene SoCs
by Kristian Vättö on October 28, 2011 4:30 PM EST- Posted in
- Datacenter
- CPUs
- IT Computing
- AppliedMicro
- X-Gene
- SoCs
AppliedMicro has released specifications of their upcoming X-Gene SoC (Server-on-a-Chip this time, not System).
| AppliedMicro X-Gene Specifications | |
| Architecture | ARMv8 |
| Cores | From 2 to up to 128 |
| Frequency | Up to 3GHz |
| Process | TSMC 40/28nm |
| Power Usage | Up to 2W per core |
Above is a simple table showing the key specifications. ARMv8 is ARM's brand new architecture, which was announced on Thursday. ARMv8 brings 64-bit addressing to ARM architecture, which makes ARM a more attractive solution for server market. X-Gene is very scalable - core count ranges from two to up to 128, while the frequency is up to 3GHz (yes, even with 128 cores). AppliedMicro has chosen TSMC as the manufacturer of the SoCs and the process will be TSMC's 40nm and 28nm.
X-Gene is a SoC, meaning that key server and network components are integrated onto the same chip. This is much cleaner approach when compared to for example Intel's, where you have several independent chips, such as the CPU(s) and chipset controller. X-Gene even has an integrated 10Gbit Ethernet controller, which should be a welcome addition for enterprises with a need for high-speed networking. Support for multi-chip configurations is also present, enabled by a 100Gbit/s interface (just for comparison, Intel's QPI is good for up to 204.8Gbit/s).
The biggest advantage of X-Gene is its power efficiency. At full load, the power usage is only 2 watts per core. When idling, the power usage is one fourth, 0.5 watts per core. For the 128-core chip at 3GHz, the power usage works out to be 256W, or 64W when idling. 256W may sound like a big number but it's actually on-par with for example two Intel X5680s, which are 130W each. And that is when excluding the power used by the chipset and other components, which are integrated into X-Gene. Of course, performance is a big question mark but if AppliedMicro's tests are to believe, X-Gene is up to three times faster than Intel's Sandy Bridge based E3 Xeons when looking at similar power profile. It should be noted these numbers are based on pre-silicon projections, so a lot can change before the final products hit the market.
The scalability of X-Gene allows a broad suite of market-end applications. The low-end chips with only a couple of cores are suitable for more consumer-friendly devices like NASs and routers - whereas the chips with higher core count are ideal for more complex setups, such as data centers. The first samples of X-Gene are expected in the second half of 2012.
Stay tuned for a more thorough analysis of this announcement!
Facebook
Twitter
RSS
22 Comments
View All Comments
DesktopMan - Friday, October 28, 2011 - link
And so it begins...twotwotwo - Friday, October 28, 2011 - link
I like the concept, but something has to change for this to take off. There are already low-power Xeons at at least one high-density many-node Atom solution (Sea Micro's SM10000) but they haven't taken over.One difficulty for ARM is that servers have lots of other power-hungry and pricey components so a CPU can't radically boost efficiency on its own. And instruction sets matter, though anything Linux is pretty portable and Windows' kernel has been ported. And Intel ain't giving ground here without a fight.
In favor of ARM: lots of work is being done on ARM speed. Datacenter workloads adapt better than most to many cores or many servers. Energy prices could spike in some places/at some times -- picture a CO2 tax, a move away from nuclear power, geopolitical trouble getting fuel, or a combination of those. And not only electrical costs but density favors low-power servers. If ARM chips sell at lower profit margins than x86 ones, that could be key, too.
I suspect there's plenty of "fat" in the datacenter -- servers that are at low CPU utilization and always will be because they're I/O or RAM-bound. Those might be candidates for ARM chips. Finally, ARM doesn't actually have to take over the DC to be important. If it starts making inroads in particular niches or for a small percent of the market, Intel's gonna feel threatened and do something exciting in response. :)
gevorg - Friday, October 28, 2011 - link
This is why its hard for Intel to compete with ARM:"Up to 2W per core"
Polymerabbit - Saturday, October 29, 2011 - link
Mobile Broadwell processors should be able to achieve that. Keep in mind that this stuff is set to be released in 2014.silverblue - Saturday, October 29, 2011 - link
Yes, but 0.5W idle? I know there's a lot of cores, but can't they be shut off entirely to conserve power even further?Soulkeeper - Friday, October 28, 2011 - link
16 or 32 cores targeted at me would be good :)iwod - Friday, October 28, 2011 - link
How about just give me 2 Core + 1G Ethernet and a Raid Controller for NAS.With ARMv8 this should be 4 to 5 times faster then today's Marvell Kirkwood NAS CPU.
mlcloud - Friday, October 28, 2011 - link
Currently using a dockstar for my file-serving and torrenting needs. Something with a bit more punch behind it in a low power envelope would be highly welcome.zanon - Saturday, October 29, 2011 - link
You write that:>Of course, performance is a big question mark but if AppliedMicro's tests are to [be believed] (little typo), X-Gene is up to three times faster than Intel's Sandy Bridge based E3 Xeons when looking at similar power profile.
That's pretty meaningless as written though. Even on a server, not everything can be parallelized to an unlimited degree. Per core performance can still matter, even without touching how well it works for int vs fp vs highly vectorized loads. This sounds like an exciting announcement, but I'm inclined to take these marketing claims with a grain or ten of salt until hard tests start popping up. It will no doubt push forward what can be done by ARM though, and any new sources of competition are always great to see.
Kristian Vättö - Saturday, October 29, 2011 - link
All of these numbers should definitely be taken with grain of salt. Like I mentioned, they are all based on pre-silicon projections (and given by APM, so no unbiased source) and it will be a year before we see the first samples.