Core 1/3 Smaller and Performance 10% Lower?
The first question that comes to mind is how important is L3 cache for CPUs on K10 (Stars) microarchitecture. The answer to this specific question will determine the market future of the new Athlon II X4 processors. On the one hand, Athlon II X2 CPUs showed us that the absence of shared cache-memory in the processor cores doesn’t lead to any dramatic performance drops. But on the other hand, it is exactly the shared L3 cache that allows minimizing the latencies during data exchange between the processor cores, so its absence in the new CPUs may make them less attractive. Moreover, in case of Athlon II X2, they partially replaced the eliminated L3 cache memory with a larger L2 cache. However, in Athlon II X4 there is nothing like that. Just like in Phenom II X4, their L2 cache is 512 KB for each core.
To dispel all doubts right away we decided to test several quad-core CPUs on Stars (K10) microarchitecture with the same clock frequency, which are equipped with L3 cache memory of different size. Here is the list of processor cores participating in this test:
- Deneb with 6 MB L3 cache. Classical solution used in Phenom II X4 900 processor series;
- Deneb with 4 MB L3 cache. CPUs like that were manufactured some time ago within Phenom II X4 800 series, but have been already discontinued;
- Agena with 2 MB L3 cache. This old 65 nm core was used in Phenom X4 9000 processors. Besides smaller cache, it also has a number of microarchitectural differences from contemporary cores manufactured with 45 nm process;
- Propus without any L3 cache. These are the new Athlon II X4 CPUs.
All four processors participating in this test session worked at the same clock frequency of 2.8 GHz.
Overall, I can say that L3 cache can have a pretty serious effect on the performance of a CPU based on K10 (Stars) microarchitecture. On average, the performance of a Phenom II X4 with a 6 MB L3 cache is about 10% higher than that of the new Athlon II X4 without any L3 cache working at the same clock frequency. However, as always, cache is not so important for a number of applications, while in others its value is hard to overestimate. For example, in final rendering tasks or video and audio content transcoding tasks Athlon II X4 works just as fast as Phenom II X4 with the same clock speed. But gamers, for instance, should better go with a CPU that has an L3 cache on it, because they can deliver much higher fps rate. Moreover, Athlon II X4 is even slower that the old Phenom X4 in games, even despite all microarchitectural improvements in the 45 nm core, including enhanced branch prediction unit and reduced cache-memory latency.
However, no one is suggesting that we should consider Athlon II X4 as an alternative to Phenom II X4 from a higher market segment. Keeping in mind that Propus processors have not only 1/3 smaller semiconductor die than Deneb, but also cost way less, we can easily forgive the performance lag we see here. Especially, if we assume that they should overclock just as good as their senior counterparts. But let’s not get ahead of our story and first see how well the new Athlon II X4 perform against the other processors from the same price range.