L3 Cache and Memory Subsystem
Touting the Gulftown as the highest-performance CPU available today, Intel puts an emphasis on two of its features: the increased number of computing cores and the larger amount of cache. However, it is clear that there are not so many applications capable of loading six CPU cores with work. These are mostly 3D modeling and digital content processing/creation software. For more popular applications, the 12MB L3 cache is going to be the key performance-boosting factor. Thanks to it, the new CPU is going to be faster at old tasks which are not optimized for multithreaded execution. The L3 cache is common for all the cores and can be monopolized by one or several cores depending on load.
A simple increase in the amount of CPU cache is always accompanied with some negative effects. Intel’s engineers did not change the logical organization of the L3 cache, so it is still 16-way associative. Considering the increased size and the arbitration among more CPU cores, the cache latency has increased by 33%.
The second factor that may have a negative effect on performance is that the uncore part, including the L3 cache and the memory controller, has a reduced frequency in the Gulftown. Intel tried this with the Lynnfield series and achieved a considerable reduction in power consumption by reducing the frequency and voltage of the L3 cache. The same goes for the new CPU. Its memory subsystem performance has been sacrificed for the sake of the two extra computing cores. Otherwise, the six-core Core i7-980X would not fit within the 130-watt TDP specified for LGA1366 processors.
Summing up the cache memory specs of the senior Gulftown, Bloomfield and Lynnfield processors, we get an ambiguous picture.
Clearly, the Gulftown is inferior to its predecessor in cache and system memory performance. The difference can be measured with Everest Cache & Memory Benchmark (I used DDR3-1600 SDRAM with 9-9-9-24 timings).
Core i7-980X (Gulftown)
Core i7-975 (Bloomfield)
The difference in practical cache performance is obvious: the Bloomfield is 33% better than the Gulftown in the speed of reading from the L3 cache and 25% better in terms of latency. The new CPU is also inferior when it comes to working with system memory. The practical bandwidth and latency of the six-core CPU is 15-20% worse than those of its four-core predecessor that seemingly has the same 3-channel DDR3-SDRAM controller.
Thus, despite the extra computing cores and the larger cache, the Core i7-980X may be slower than the Core i7-975 in practical applications due to objective reasons. This explains the rather low model number of the new product, by the way. The new Gulftown is not better than the Bloomfield from every aspect. It is weaker in some important respects.