Synthetic & Scientific Benchmarks
Starting with the single core portion of GeekBench 3, we see the first evidence of performance issues with the 5960X. Since it only has a clock speed of 3.0-3.5 GHz, it has a lower GeekBench score than either the 5930K or 4960X. What is interesting is that the 5930K outperforms the 4960X even though it's maximum turbo frequency is .4 GHz lower. This is a good sign for the Haswell-E architecture as a whole and suggests that the 5930K may outperform both the 5960X and 4960X in single-threaded applications.
One thing we want to point out is the low memory score on both the 5960X and 5930K. These CPUs use the new DDR4 memory running at 2133MHz versus the DDR3 memory running at 1600MHz we used with the 4960X. One benchmark is obviously not conclusive, but this does suggest that DDR4 memory might not have much of a performance advantage over DDR3 quite yet.
Moving on to the multi core portion of GeekBench 3, we see where the 5960X really shines. With eight cores, it easily outperforms both the 5930K and 4960X. Interestingly, the 5930K performs about the same as the 4960X in this test even though its single core performance was better.
Unlike the single core benchmark results, the memory score with DDR4 on the Haswell-E CPUs is actually higher than the 4960X.
CineBench R15 is a great rendering benchmark based on MAXON's CINEMA 4D software that measures single and multi core performance as well as OpenGL graphics performance. In this benchmark, all three CPUs get effectively the same score on the single core portion of the benchmark, but the 5960X really pulls ahead in the multi core portion.
The OpenGL portion of CineBench is supposed to measure the performance of the GPU so we didn't expect to see much difference between the three CPUs. Oddly, both of the Haswell-E CPUs outperformed the 4960X by a healthy margin. This may just be an aberration, or it may be an indicator that GPU intensive tasks might see a benefit when using a Haswell-E CPU.
Linpack is a standard benchmark in the scientific community that measures a system's floating point computing power. This benchmark can take advantage of the AVX2 instruction set that is included in Haswell-E which is clearly shown by the large increase in performance. The 5930K is anywhere from 27-75% faster than the 4960X and the 5960X is anywhere from 45-114% faster depending on the problem size.
This is amazing for the scientific community, but we want to be clear that for the average user this is not really an indicator that Haswell-E will double your performance. To find out how much of a performance increase you could expect in a variety of situations, lets move on to some real-world benchmarks.
To benchmark the performance of Photoshop, we applied a number of effects to both a 110MB and 1GB image and recorded how long it took for each effect to finish. These effects include CMYK Color Conversion, RGB Color Conversion, Ink Outlines, Dust and Scratches, Watercolor, Stained Glass, Lighting Effects, Mosaic Tiles, Extrude, Smart Blur, UnderPainting, Palette Knife, and Sponge.
The 110MB image shows roughly a 5% drop in performance with the Haswell-E CPUs compared to the 4960X. The 1GB image also shows a drop in performance, but the 5960X is a bit better than the 5930K. What this means is that if you use Photoshop a lot and already have a 4960X CPU, you probably should not upgrade to a Haswell-E CPU.
Unlike Photoshop, Premiere Pro appears to like the new Haswell-E CPUs. The parts of the benchmark that uses GPU acceleration doesn't show much difference between the three CPUs, but the portion that is CPU only shows almost a 30% performance increase with the 5960X compared the the 4960X. The 5930K is also better than the 4960X, but only by about 5%.
The Cadalyst benchmark for AutoCAD 2015 is interesting because it shows a slight performance increase with Haswell-E on the CPU and 2D Graphics tests, but a larger performance increase in the 3D Graphics portion. This increase in 3D performance is reminiscent of the OpenGL performance in CineBench we saw earlier and is something we did not expect to see.
Handbrake is an open source video transcoder that reports performance in the number of frames it is able to process per second. As you can see, the 5930K is almost identical in performance to the 4960X. The additional cores on the 5960X, on the other hand, give it a nice ~10% performance boost over the 4960X.
Unlike Handbrake, Lame is not multithreaded so the number of CPU cores you have does not affect performance. Because of this, both the 5930K and 5960X (both of which have a lower frequency than the 4960X) show about a 6-9% drop in relative performance.