Intel's 22-core Broadwell chips will speed up cloud services

Intel packed more cores into the chips in response to growing demand for cloud and mobile services. The greater number of cores will allow cloud providers to stream more movies or applications from each individual server.

The chips, based on Intel's Broadwell microarchitecture, will also go into workstations. Paired with a high-end graphics processor, workstation users will be able to develop virtual reality applications and edit 4K video. Those workstations will likely use the quad- and six-core variants of the new chips.

The Xeon E5-2600 v4 lineup includes 27 chips and boasts many improvement that make the chips faster than their predecessors. Each core is roughly 5 percent faster than those based on the prior, Haswell architecture.

Based on standard SAP benchmarks for the Linux OS, Dell found the new chips were 28 percent faster than their Haswell equivalents.

But cooling could be an issue for chips that have so many cores. Intel dialed down the frequency of the monster 22-core Xeon E5-2699 v4 to 2.2GHz, which draws about 145 watts of power. It has 55MB of cache, up from 44MB of cache in its fastest predecessor. The 22-core chip is priced at $4,115.

The new chips are primarily targeted at cloud service providers. Applications are moving into public, private and hybrid clouds, and the chips are tuned to speed up and secure hosted applications, said Matt Lane, director of Xeon marketing at Intel.

In 2015, most of Intel's top seven cloud customers upgraded to servers with more cores, Lane said. Cloud providers want servers with more computing power, in order to rent out more virtual servers and pack more cloud services onto each one, Lane said.

Cloud providers like Google and Facebook build vast data centers based on Intel chips, but they're also considering alternative architectures like ARM and IBM Power. Qualcomm is making a 24-core ARM server chip that reportedly has support from Google.

Intel is sitting pretty today, with some 90 percent of the server market. ARM and Power have virtually no market share, but they're getting enough attention for Intel to be concerned and take action to retain customers.

Intel will customize the new Xeons for larger customers, Lane said.

In addition to cloud providers, the new chips are targeted at companies running databases, analytics and ERP (enterprise resource planning) applications.

The chips plug into sockets in existing servers currently using the Xeon E5-2600 v3 chips based on Haswell, which started shipping last year. As a result, Dell is providing incremental upgrades to existing servers like its PowerEdge R730, which received a cooling system upgrade to handle new chips with more cores. Hewlett Packard Enterprise, Lenovo and others are also expected to announce servers with the new parts.

Some of the chip improvements are reflected in the finer details. Some Xeon chips will support a version of DRAM DIMMs in which memory chips are stacked on top of each other in a 3D format. That memory type will be available in the future.

The Xeons also have new instructions that speed up security tasks like encryption and decryption. Lane said security tasks are being processed 70 percent faster.

Another interesting feature is the Resource Director Technology, which helps speed up virtualization. The technology has a cache allocation feature, which can give guaranteed cache space for high-priority virtual machines. It also helps in software-defined networking and NFV (network-function virtualization) environments, where high-priority data packets can be moved into the cache ahead of others.

An add-on card available at a later date will allow the implementation of Intel's proprietary OmniPath fabric interconnect, which boosts data transfer speeds between storage, memory, processors and other components. Intel has not shared details on the underlying technologies in OmniPath, which is targeted mostly at high-performance computing.

The chip's integrated networking, storage and I/O support are almost identical to the predecessor Haswell server chips. That's a trade-off Intel had to make to keep the new chips compatible with existing sockets in Haswell-based servers so upgrades are less costly for customers.