Nehalem: Xeon Gets Core i7 Upgrade
Posted on: 03/29/2009 05:00 AM

SiSoftware SANDRA 2009 SP2
SiSoftware, founded in 1995, is one of the leading providers of computer analysis, diagnostic and benchmarking software. The flagship product, known as "SANDRA", was launched in 1997 and has become one of the most widely used products in its field. SANDRA is used by almost 500 world-wide IT publications, magazines, review sites to analyse the performance of today's computers. Over 5,000 on-line reviews of computer hardware that use SANDRA are cataloged on the company's website alone.
SiSoft's SANDRA is a good way to paint a picture of overall system performance. It is important to note that because SANDRA is a suite of synthetic tests, the numbers don't always correlate directly to "real world" performance in "real world" applications. That being said, it is a good stress test, and it does allow us some insight into peak theoretical performance.

We'll start with the CPU tests...

Considering the clock-speed advantage of the Harpertown system, Nehalem's margin of victory is pretty impressive. While the basic cores of the two processors are the same, optimizations in the pipelines, cache, etc seem to make a pretty big difference.

SANDRA has a relatively new test in the CPU suite that tests multi-core efficiency. The test works by shuffling a variable number of different sized blocks in and out of the L caches on the processor(s). Some test fit into a single processor's caches while others spill out into both processors' caches.

The Nehalem is pretty impressive here until the number and size of the block increases, but having a larger cache really gives Harpertown an advantage as Nehalem falls off.

I mentioned the FB-DIMM/DDR3 thing earlier, so let's see what light SANDRA can shed on the subject.

What is there to say really?! Wow! QPI rocks... or something. Things are pretty even between the systems until the block size spills out of the cache and into system memory. Nehalem just takes over and completely dominates in the memory bandwidth department. I haven't seen these kinds of numbers since, well, Opteron.

Let's see how latency is affected by the switch to DDR3.

Again, once the block size spills out into main memory, Nehalem simply runs away with this test. If the synthetic numbers are any indication of real-world performance, I think we're in for some record setting performance numbers.

Printed from (,3.html)