Intel Announces New High-Efficiency Atom C2000, Demoes Silicon Photonic Interconnect
Based on what we know of Bay Trail/Avoton, we expect it will offer significantly improved performance over existing S1200 Atom servers, thanks to multiple microarchitechtural improvements and the integration of a quad-core processor rather than dual-core + Hyper-Threading. What's significant about the microserver part is that Intel is claiming these new chips will fit into the same 6W power envelope as the lower-performing in-order hardware.
Intel's claims here should be taken with a bit of explanation. The "Up to 7x Faster" is impressive, but the Intel C2750 is being compared against an Atom S1260. The S1260 is a dual-core / quad-thread processor with 8GB of RAM and a 1GbE ethernet link. The C2750 is an eight-core processor with 32GB of RAM and a 10GbE linkage. Granted, the 32GB of RAM and faster link make sense in terms of keeping the CPU fed with data, but we're mixing apples and oranges comparing dual cores against eight.
Of more interest is the performance-per watt claim. While these figures are drawn from particular applications, they do appear to show that the octa-core variants of the C2000 will be capable of delivering far more performance than the current Atom, while remaining in the same 20W per server power consumption level. That figure appears to include both an SSD, HDD, and 8-16GB of RAM rather than simply focusing on the SoC itself.
Photonic Interconnects Coming Soon
The other major new technology up for discussion is the silicon photonics interconnects that Intel is working on in conjunction with Corning. To date, high speed interconnects between systems have relied on copper wire, but significant problems with continued scaling have led Intel, IBM, and other companies to search for alternatives. Copper is relatively cheap and plentiful, but increasing bandwidth over wire inevitably requires substantial tradeoffs in link distance, cable size/weight, power consumption, and routing. One cable between two systems is not a major issue, 500 cables between server racks is an enormous airflow impediment and takes up a great deal of space.
Hence, silicon photonics. This is another technology that isn't shipping quite yet, but Intel showed off the hardware in action at its Hyperscale Datacenter event today. Each link is theoretically capable of up to 1.6 terabits per second or about 200GB/s of bandwidth in total. Each cable is made up of 64 fibers, with each fiber running at 25Gbps. Interconnect speed at 300 meters is estimated as 25Gbps in total, which is a sharp falloff from the peak, but also represents a distance of nearly a fifth of a mile. Most racks aren't separated by anything like that much space.
The long-term impact of silicon photonics will be felt first in datacenters and HPC buildouts, where they'll be used to significantly reduce power consumption and rack-to-rack communication latency. Long-term, we should see them integrated into microprocessors and SoCs. The goal of both these technologies is to allow for data centers that are more flexible, smaller, and can be quickly provisioned or taken offline depending on shifting needs within an organization.
But there is a political angle to the announcements as well. Intel has been stepping up its own drum banging as interest in ARM servers and non-x86 datacenter deployments has grown, fueled by companies like Calxeda and X-Gene. With AMD jumping into the low-end microserver market with both an ARMv8 64-bit core and its own Kabini-based hardware, Intel is acting to head off a potential threat to a market it's determined to defend.
Don't miss the financial angle. With x86 consumer sales taking a beating, Intel's high-margin server business has become even more important to its continued financial health. The company is determined not to allow a nascent ARM threat to steal sales from a vital component of its business.