TheOverclocker Issue 17 | Page 12

(even though they are supposedly measured differently) but most of all it’s the exact same Asetek cooler used for both. Performance Analysis If you were expecting to be blown away, well you will not be. Very few programs that we are concerned with benefit from the additional cache that this CPU has and the rest are benefitting from the four extra threads if anything. Our testing methodology much like it is in the FX feature is slightly different from the norm. We tested using a two-way Crossfire system, which would allow high end CPUs to flex their muscles or in a more practical sense show their worth. It is very easy to reach GPU saturation in a single GPU configuration especially for the overclockers that are only as brave as water-cooling allows them to be. The differences in CPUs are just simply easier to see when there isn’t a massive bottleneck courtesy of the graphics card and given that all X79 boards on the market support at least three graphics cards in tandem for now, why not test with 2 cards anyway. One thing worthy of noting as well is that, we tried dual, triple and quad channel configurations but there was no difference in performance. The variance in the numbers was within the margin of error and we can only conclude that while synthetic benchmarks (aren’t they all? -Ed! ) show sizeable gains in available memory bandwidth with each populated channel, in practice the difference just isn’t’ there. ArmAII, is slightly slower on the 3960X than on the 2600K, however it’s only 2fps and this is purely because of the 100MHz frequency deficit of this CPU. As such, it’s nothing that a small multiplier bump cannot remedy. Lost planet 2 however swings the other way in favour of the 3960X. Resident Evil 5 however showed the biggest gains scoring nearly 20fps higher than the 2600K and over 56fps higher than AMD’s best showing. A single and relatively old game is no reason to buy a CPU, but if ever there were tangible benefits in gaming from the purchase of this CPU, this benchmark has shown them. 3DMark03 seemed to love this CPU, doubt if it’s thread related, but most likely related to the vast quantities of cache or simply 12 The OverClocker Issue 17 | 2011 the 4-way IMC working it’s magic, whatever it is, this platform will rule 3DMark03 provided you can find a CPU that can clock sufficiently high. The rest of the synthetic tests save for 3DMark05 are benefitting from the additional threads of the CPU, delivering scorching performance all around. 3DMark05 was a peculiar one, suffering from the lower clock of the 3960X. Do keep in mind that while some of the results don’t show a massive gain over a 2600, it’s important to realize that this is only the overall score and not the individual CPU test scores. The 3960X scores massively here and benchmarks like 3DMark11 and 3DMark06 for different reasons undersell that difference. The Vantage result is probably the most telling of all the 3D tests as the numbers are very impressive especially as the clock speed increases. Out-Take We observed something while working with the 3960X that we didn’t expect or at least had not experienced with the LGA 1155 CPUs. It turns out that if you disable cores, you actually decrease the total available L3 cache. An obvious assertion one would make right. Well, not exactly. See on the 2500/2600/2700 CPUs, disabling the cores and even leaving two functional doesn’t necessarily rob you of performance in largely single threaded benchmarks like 3DMark05 for example. However, we disabled all but two cores on the 3960X for heat purposes but found that the 3DMark03 scores dropped dramatically, in fact in the region of 5,000 points for 3DMark03. CPU-Z reported the same cache amount as before at 15MB, but when reading the literature provided by Intel it turns out that each core can only ever have access to 2.5MB of L3 cache exclusively. So with only two cores enabled only 5MB of L3 cache was available in total, less than that on a 2500K even. We are not equipped with the right tools yet to do cache latency tests, scaling and controlled workloads but suffice to say, this is the only logical conclusion we can reach right now. We may be completely of the mark, but we did not observe this difference in scores when disabling half the cores on the