Jason Waxman, General Manager of Intel’s Cloud Infrastructure Group (Fig. 1), revealed Intel’s 22-nm, 8-core C2000 Atom system-on-chip (SoC) based on the Silvermont architecture (see Intel’s Silvermont Architecture Targets Low-Power Applications). It targets server applications handling up to 64 Gbytes of ECC DDR3 memory. It is also destined for storage and comms applications.
The C2000 (Fig. 2), code named Avoton and Rangeley, is designed for high efficiency compared to the high performance Intel Xeon and Intel Phi (see Battle Of The Supercomputing Nodes). The SoC also incorporates Ethernet, USB, SATA and PCI Express Gen 2 controllers.
Eight cores let the chip do a lot more heavy lifting. Diane Bryant, senior vice president and general manager of the Datacenter and Connected Systems Group, talked about Intel’s Rack Scale Architecture (RSA) based on the Open Network Platform reference design. It essentially splits the conventional server components like power supply, networking, storage and compute into their respective components that are then modularized and shared. For example, the power supplies for a compute rack would provide DC power to the compute rack that would contain processor modules and network modules. The processor modules might have a single Xeon or a bank of C2000 Atom cards.
The modular approach is designed to reduce costs by sharing modules while providing flexibility in configuration. The Xeon card could be swapped for the block of Atom cards. The choice would be based on the power/performance required for the application.
Packing lots of low power processors in a cluster has been done in the past. For example, the original version of AMD’s SeaMicro SM1000 (see 10U Rack Packs 512 Atoms) used 256 dual core Atom processors. They also used a custom ASIC to implement a high speed communication fabric. Intel’s approach would use the on-chip Ethernet support connected to a switch module.
Calxeda takes a similar approach with quad-core Arm SoCs. Its ECX-1000 Series uses a 10 Gbit/s link fabric. The SoCs have the fabric support built in.
Intel’s use of Ethernet has more overhead but is more flexible from a chip standpoint since the same chip can be used in a server cluster or in a standalone application. Intel’s 64-bit chip has the edge when it comes to computing compared to the 32-bit Arm SoCs but Arm’s 64-bit platforms (see Delivering 64-Bit Arm Platforms) will be coming online soon which will make this discussion more interesting.
The low power and high integration of the C2000 Atom makes it interesting for high performance embedded applications. So far Intel has been successful with this approach.