Intel's Spectacular New Server CPU

Last week Intel tossed a bombshell at the server market. It came in the form of the Nehalem-EX, an eight-core powerhouse that could redefine both the performance and economics of the multiprocessor sector.

Intel has done very well with the Nehalem microarchitecture, first with the Core i7, then later featuring it in the Core i3 and i5 processors. There are also the six-core Core i7-980X and Xeon 3600/5600 series CPUs based on the Gulftown iteration of the Nehalem core, but the Nehalem-EX is something different.

Nehalem-EX, or officially the Xeon 7500 series, is the next-generation Intel CPU for the "expandable" server market, which are those systems with four or more processor sockets. This environment goes far beyond traditional, relatively small two- or four-way servers and can incorporate hundreds of interconnected multicore processors all working in concert on mission-critical tasks. The Nehalem-EX is certainly ready to compete in this area, as it can support up to 256 chips per server and offers an unprecedented feature set for an x86 processor.

The Nehalem-EXposed

The Xeon 7500 utilizes many of the common strategies Intel employs when launching a new and improved product line. Increasing the number of cores is a popular one, and the Nehalem-EX jumps up to a full eight-core design (in addition to quad- and six-core models). This is not one of those "two quad cores on the same die" kludges, but a true octo-core designed from the ground up. This means shared resources, greater efficiencies, no latency-filled CPU interconnects, and lower thermals and TDP ratings compared to a dual-on-die design.

Having eight cores obviously ramps up multithreading capabilities, and with Intel's Hyper-Threading Technology doubling this to a maximum of 16 simultaneous threads per CPU, the Xeon 7500 should be able to fly through high-end, parallelized code -- 64 threads of it per four-chip or 128 threads per eight-chip platform. Virtually all server code at this level is multithreaded, so Hyper-Threading is an important feature, especially since it was not available on previous Core-based Xeons. The Xeon 7500 series also supports Intel Virtualization Technology and can run either 32- or 64-bit x86 software.

The 7500 features a built-in memory controller that utilizes DDR3 memory in a quad-channel configuration at up to 1066MHz clock speed. The new processors also include Intel Turbo Boost technology, which combined with scalable memory interconnects provides up to eight times the memory bandwidth of previous Xeon platforms.

Memory capacity has also increased. In a four-processor configuration, the Nehalem-EX and its scalable memory buffer can use up to 16GB DDR3 DIMMs and support 16 DIMMs per CPU. This means it can handle up to 1TB of system memory in a four-socket system (2TB in an eight). Due to its quad QuickPath Interfaces, the Xeon 7500 can scale up to an eight-socket configuration, with the ability to scale even higher (to 256 sockets) using third-party node controllers.

Performance specifications vary based on the model number, with the eight-core Xeon X7550 sporting 18MB of L3 cache and the Xeon L7555 running at a QPI speed of 5.86 GT/second. The top-end Xeon X7560 ramps this up to a full 24MB of L3 and a QPI speed of 6.4 GT/second. These are both common methods of differentiating the various Nehalem-based processors, and can have a real impact on performance levels, especially memory bandwidth.

All of this comes at a cost, as Intel obviously could not fit all these improvements into an existing socket. The Nehalem-EX brings with it the new LGA 1567 package, which bypasses the LGA 1366 socket used by most Nehalem-based processors and replaces the Socket 604 of previous Xeons. This eliminates some basic upgrade paths, but also ensures that the Nehalem-EX is running on a platform tailored to its specific needs.

The Nehalem-EX is built using 45-nanometer process technology, rather than the sexier 32nm design of the six-core Westmere-based workstation and small-server Xeon models. The larger die size results in lower clock speeds for the Xeon 7500 models, from 1.86GHz to 2.66GHz, compared to well over 3GHz for the Xeon Westmere models. This does keep the TDP specs down, and these top out at 130W for the 2.26GHz Xeon X7560.

Reliability, Availability and Serviceability

While the performance specifications are certainly important, the Nehalem-EX is a lot more than just an ultra-fast, eight-core powerhouse. In fact, its most significant improvement is in closing the gap between x86 server solutions and their RISC counterparts. In previous Xeon designs, no matter how powerful, large-scale implementations were not a viable RISC alternative because these x86 processors lacked some integral features.

Error detection and recovery is an important one, and the Nehalem-EX has made significant advances in this area. Previously, you would require a RISC system to ensure that hardware failures could be seamlessly located, blocked off and then hot-swapped out, all in real time and without crashes or needing to turn the system off. This is referred to as Machine Check Architecture (MCA) Recovery, and it is the first time we're seeing it in an x86-based platform.

Intel has also implemented features like QPI self-healing and SMI lane failover to allow the system to detect problems and perform the necessary modifications to keep the system up. These are not easy to implement; keeping the operating system running not only during an error, but also through the entire diagnostic and repair phases as well, is no mean feat. There are other similar features such as the ability to hot-swap or add both memory modules and processors, allowing performance upgrades on the fly.

AMD: 12 Cores to the Rescue?

AMD also has its own new contender in the multiprocessor server arena, dubbed the Magny-Cours or Opteron 6000 series, and it has 12 cores to boot. It's not quite as revolutionary a product as the Nehalem-EX, since the Magny-Cours is just two six-core processors bolted on the same die, but it's still a dozen processor cores for excellent 12-threaded performance.

The two dual-channel DDR3 memory controllers are roughly equivalent to the quad-channel of the Nehalem-EX, and AMD's DDR3 is clocked higher at 1333MHz. But it pays a price with a 2.2GHz clock speed and lack of shared resources between the two six-core CPUs, so performance suffers. Once again, AMD is playing its well-worn value card against a higher-priced Intel product.

The Nehalem-EX is not only a powerful, feature-rich processor, it is also coming in at a very opportune time. Current economic conditions have led companies to keep older servers much longer than anticipated and this will lead to forced upgrades throughout the corporate world. With the eight-core Xeon 7500 series looking like a very attractive, reliable and cost-effective upgrade compared to a new RISC machine, this could spur a perfect storm for Intel's server division.