Intel's Xeon "Nehalem" Processor. The Empire Strikes Back.

Intel is set to officially launch their new line of Xeon processors this month. This release is the latest "tock" in Intel's tick-tock development model, and is based on a new microarchitecture code-named Nehalem. Intel has been following the tick-tock model (refine existing architecture one year/create new architecture the next year) for the last few years so technically this release shouldn't be big news, but based on all the buzz being generated, it looks like this going to be a really loud tock. Both Sun and RedHat have specifically mention Nehalem support in their latest release: OpenSolaris 2008.11 and RHEL 5.3. Citrix and VMware have also voiced support.

Image courtesy of Apple

Intel vs AMD

Since the launch of AMD's Opteron line in 2003, Intel has struggled to regain performance leadership in the server CPU market. The two have fought back and forth feverishly, and while Intel has still managed to (controversially) dominate in terms of market share and profitability, the performance question has been a topic of much debate. With the launch of the Nehalem microarchitecture Intel is seeking to definitively reclaim its crown.

Features

To the careful observer many the Nehalem's "new" features don't seem very new at all. Most of them are features that AMD implemented several years ago.

Multi-core, Single die - The Nehalem Xeons will sport 4 (Xeon EP) or 8 (Xeon EX) cores, all on a single die. Previous quad-core Xeon processors had two pair of cores on separate dies, which meant that some cached data had to travel outside the processor to get from core to core. The new architecture is more efficient. AMD has being using this approach all along.

Integrated Memory Controller - Rather than connecting system memory to a processor through a separate I/O controller, the Nehalem features an integrated memory controller. This means faster access to data stored in memory, and significantly reduced latency. This was one of the AMD's major features when it launched the Opteron...six years ago.

QuickPath Interconnect - In multiprocessor configurations, QuickPath acts as a high-speed interconnect between processors (and each processors memory bank). With the introduction of QuickPath, Xeons can now take advantage of Non-Uniform Memory Access which allows them to scale better as the number of processors increases. QuickPath also acts as the transport for connecting the processors to the motheboard's Southtbridge or IO Hub. AMD calls this HyperTransport.

Hyper-threading - Now here is something they didn't copy from AMD; nevertheless it isn't exactly new either. Hyper-threading allows two threads to run simultaneously on each core. So an 4-core Xeon-EP presents 8 virtual cores to an operating system. This is an Intel technology that made its debut with the Pentium 4, but was since shelved. Looks like it is back and performing better than ever.

Virtulaization - The Nehalem introduces two important virtualization features, Extended Page Tables, which significanlty reduces the overhead involved in virtualizing memory access. (AMD calls this nested page tables and has support in its newer processors) and VT-d, which allows virtual machines to be given direct access to physical devices ( AMD calls this IOMMU and expects to have support by 2H 2009). Technically this is actually a motherboard chipset feature.

The Numbers

Now that Intel and AMD seem to be reaching feature parity in terms of their overall chip architectures we will probably see even more fierce competition. But if Intel is basically just playing catch up to AMD in terms of architecture, what is the big deal? In one word, performance.

While Nehalem desktop chip (Core i7) performance improvements are in the 15-20% range, on the server it is whole different story. This HP/SAP benchmark shows the Xeon Nehalem beating its predecessor by about 100%.

Here is what Red Hat had to say:

"In internal testing, the Red Hat Engineering Performance Group has measured exceptional gains with the new Nehalem processors, with unaudited results showing gains of 1.7x for commercial applications and gains up to 3.5x for high-performance technical computing applications compared to the previous generation of Intel processors."

Apple got the jump on the big server vendors and announced on Tuesday that it is shipping Nehalem Xeons in the Mac Pro workstation line. Here are their numbers:

"The result is fast access to cache data and greater application performance. Combine that with the other technological advances and you get a Mac Pro that’s up to 1.9x faster than the previous generation."

"The integrated memory controller, along with fast 1066MHz DDR3 ECC SDRAM, also gives Mac Pro up to a 2.4x increase in memory bandwidth over previous generations."

AMD should benefit for some of the software optimization being made for the Nehalem, but come the end of March it looks like Intel will have a clear all-around lead for the first time in a long time.

The Empire has struck back, and it only has one question for AMD. Who's your daddy now?