Intel Pentium 4 6xx and 3.73EE: Favoring Features Over Performance

Twice the Cache - 17% Higher Latency

Both the Pentium 4 6xx and the new Extreme Edition share the same core, meaning they also have the same L2 cache.  When Intel first launched Prescott we noticed that in the move to the new architecture that cache latencies went up tremendously.  The increase in cache latencies was to be expected, as one tradeoff of a larger cache is that it takings longer to find and access data.  So when we heard that Intel was moving to a 2MB L2 cache with the 6xx series, we wondered how much slower the cache would get.

First we wanted to confirm that L1 cache latencies stayed the same, and they did at 4 cycles for the new Prescott 2M based core:

Next up, was L2 cache latency.  In our review of the Pentium M processor on the desktop we discovered that its 10 cycle L2 cache was responsible for its solid performance in non "media rich" applications (e.g. office applications, OS performance).  The original Prescott had a 23 cycle L2 cache, and with a 2MB cache the latency has gone up to 27 cycles:

While we're talking about "only" 4 cycles, at 3.6GHz that's 17% longer to access data from L2 cache.  Given Prescott's extremely lengthy pipeline, a 17% increase in L2 cache latency is not going to help minimize the downsides of such a long pipeline.  Also keep in mind that the only architectural change here is a larger L2 cache, so none of the normal tricks to help hide memory latencies are expanded upon in the new Pentium 4. 

What Intel is counting on is that the increase in hit rate provided by a 100% larger cache will outshine the 17% longer access to L2 cache.  Did Intel make the right bet?  In order to find out we took the new Pentium 4 660 (3.6GHz - 2MB L2) and compared it to the old Pentium 4 560 (3.6GHz - 1MB L2), with all other variables the same, let's see how much of an impact the extra megabyte of cache has in the real world.

In the business category, we see the added cache paying off a little. SYSMark shows good improvement in the document creation portion of its tests, while the Business Winstone makes some very good gains. Worldbench shows web browsing with Mozilla to have improved a good bit while our compression test and the ACDSee test show a loss in performance. These losses generally indicate areas where the test is more dependant on latency than cache hit rate. On the content creation side, adding Windows Media Encoder to the Mozilla test improves performance more than the individual Mozilla test. This is likely due to the fact that the large cache keeps Mozilla's data from being kicked out while Windows Media Encoder is working.

On the gaming front, Doom 3 is the only test we saw with any performance improvement. And the only other application to show a significant performance gain is Maya with more than a 43% gain. The huge gain in performance under Maya is likely a result of 1MB of cache being too small to fit models in while 2MB is enough. This seems to be a case where the test is very bandwidth sensitive rather than latency sensitive. Dropping most (if not all) of the data being worked on into the L2 cache offers a program a very large boost in apparent bandwidth.

As we can see, the unfortunate truth for performance on the 600 series is that most consumer data sets can fit into a 1MB cache just fine. The added cache does seem to help with multitasking from our limited investigation of the subject. The more threads that hit memory aggressively, the better chance we have of seeing a benefit from the 2MB cache. This is because less data from each thread will be kicked out of the cache, resulting in fewer pipeline stalls.

Unfortunately, most usage models that are a good fit for the 600 series are server and workstation workloads. Streaming data (using or encoding media), games, and most other consumer applications don't have the lots of big data requirement that can really separate the performance of the 1MB and 2MB parts.

As we've provided this chart and gone through the general impact of the benchmarks on Intel's new 600 line, we won't include analysis on the pages with our benchmark data. For those who are interested in a deeper look at the numbers and performance of all 5 new parts, graphs of each benchmark are included later in this article.