For a couple of years now, I’ve been dreaming of building a low-power ARM server cluster for… something. I never really got the process very far; early ARM server SoC providers were either unwilling to respond to inquiries or completely unhelpful. Oh yea, and I never had a real reason for doing it, other than, you know, doing it. A week ago, that dream was reinvigorated by an actual use case. However, I “need” 64-bit addressing, so my drawer of odd ARM development boards I’ve accumulated during spouts of ambition over the last couple of years is useless. But, my iPhone 5s sported an ARMv8 64-bit processor, so surely there must now be some solid 64-bit SoCs out there?
With AMD’s 64-bit ARM server revolution still failing to materialize, Calxeda going belly up and Marvell, Qualcomm and the rest focusing on mobile applications, it turns out the answer is “no.” … -ish. The only real contender with an actual product to show for is AppliedMicro with their X-Gene SoC. Although the development boards sport some pretty serious price tags ($1495 for the basic and $2495 for the fun one), an 8-core 2.4 GHz SoC with 10G Ethernet, a SATA III controller and ECC RAM support touting “Xeon-level performance” surely must be able to do the trick? After all, HP is now shipping microservers with this next-generation piece of silicon (points to the marketing division for repurposing the name “cartridge.”)
As it turns out: no. AnandTech has done an excellent review, pitting the 1st generation X-Gene (the only one available) against some Atoms and low-end Xeons. The results are striking; the archaic 40 nm process with which the X-Gene 1 is produced coupled with the apparently abysmal performance means that AppliedMicro’s big bet fails to deliver in every single way – including performance per unit of power. And in a big way.
As with any benchmark, there are some details to be kept in mind. The compiler generation used in the review doesn’t have the latest in AArch64 optimizations, and this is of course only the first generation of the chip, with X-Gene 2 supposedly to be produced with a more modern 28 nm process. But, none of that really matters. As far as compiling your code and running it in production goes, the X-Gene 1 and the AArch64 environment at large seem to be so far off the mark it’s not even a hypothetical contender at this point. For now, then, it seems we’re still better off in any regard just buying mid-range Xeons and getting on with it.