Intel’s Performance Claims
In our last visit to Hillsboro, Intel provided quite a few performance estimates comparing Dempsey to previous generation products from Intel, and existing products from AMD. These numbers really served to highlight the benefit of dual core MPUs equipped with a chipset designed from the ground up for dual cores. This time around, we were shown estimates for Woodcrest systems compared to prior products and current AMD offerings. While this information should help to show both the performance improvement from Woodcrest and the competitive landscape, there are several qualifications.
First of all, these estimates are for benchmarks, where performance engineers from each company spend hours tuning and tweaking the performance to get the best score possible. The results may differ when little tuning or optimization is done. Second of all, the hardware used for these results is always top of the line, the most expensive systems out there. Most users will probably not buy the best and opt for something more economical. Lastly, AMD is in the process of releasing the new AM2 socket and MPUs across their entire product line. Socket AM2 will use DDR2 instead of DDR1, increasing the bandwidth. By all accounts, AM2 processors will have anywhere from 1-10% higher performance, although most workloads will only improve by 1-5%. Moreover, AMD may very well bump up their clock speed to 2.8GHz or even 3.0GHz if they are lucky. So with those qualifiers in mind, let’s see what Intel tells us about Woodcrest.
Figure 2 – Relative Intel System Performance
Figure 2 above compares the performance for Intel systems over three generations, according to Intel. Irwindale is a 90nm, single core, 3.8GHz MPU, with 2MB of L2 cache and an older chipset. Dempsey is basically a pair of Irwindales shrunk to 65nm in a single package, using the newer Blackford chipset. Woodcrest uses a slightly faster version of the Blackford chipset, with a much better MPU, also at 65nm.
The scaling for Dempsey is reasonable, an across the board 50-90% improvement for going to two cores. However, Woodcrest is simply amazing, it is twice as fast as the 90nm Irwindale on almost every benchmark, and on SPECjbb2005 (which is the best indicator of server performance in the sample), it is 2.75x faster. While Woodcrest is really head and shoulders above Irwindale, it is also a very noticeable improvement over Dempsey. For the workloads shown above, it is 10-50% faster, in the same process technology.
The one exception to these improvements is SPECapc using Solid Works. Based on the results above, it appears that this particular benchmark is most likely single threaded, or written in such a way that it does not benefit from more than a certain number of threads or processors. Solid Works does speed up by about 40% with Woodcrest, but that appears to be consistent with single threaded gains. Moreover, if there was any benefit from having more processors, surely Dempsey would see improvements, yet it does not. While there is a strong graphical component to SPECapc, 3DS Max scales moderately well with the newer processors, so clearly additional CPU power can be useful for the SPECapc benchmark. Why the Solid Works engine is not able to use the additional CPUs is unclear, but it certainly is a special case worth mentioning.
Figure 3 – Intel Competitive Estimates
Intel also provided a direct comparison between Paxville, Dempsey, Woodcrest and the best published Opteron systems in several benchmarks. Considering the source, all these numbers should be taken with a grain of salt. Now the really interesting question is will our real world benchmarks show the same numbers as Intel’s projections…