Benchmark Results: Sandra 2012 SP3
More so than PCMark or 3DMark, Sandra is good for extracting the potential performance from each of these CPUs.
Clearly, the six-core Sandy Bridge-E parts serve up the highest numbers in Sandra’s Arithmetic test. Ivy Bridge’s slight IPC improvements translate into a comparably slim advantage over Core i7-2700K, and both of those parts are substantially faster than the remaining three processors.
I broke out the multimedia results into two charts, since Phenom II doesn’t support AVX. Integer performance is a strong suit for AMD’s FX processor, which makes sense given its eight integer cores. Floating-point throughput isn’t as compelling. Again, that’s expected given the shared nature of the architecture’s resources devoted to floating-point math.
Leveraging AVX instructions, the eight-core FX is even able to outperform Intel’s Core i7-3770K in the integer component of this test. AMD is humbled by the floating-point portion, though, succumbing even to Intel’s Core i5-2550K. Its four modules (with four corresponding floating-point units) cannot keep up.
Ivy Bridge, meanwhile, is only marginally faster than Sandy Bridge at the same clock rate.
The clear winners here are the Core i7-3960X and 3930K, which brute-force the test using six cores.
Hardware support for AES encryption and decryption means that the Ivy Bridge, Sandy Bridge, and Bulldozer architectures can operate on data as fast as it’s delivered. Thus, the AES256 results scale according to memory bandwidth—the bottleneck in this equation. SHA hashing is a much slower process, which is dominated by the Sandy Bridge-E-based parts.
And here are the bandwidth numbers to drive our Cryptography test results home. Operating at the same DDR3-1600 data rate, Ivy Bridge doesn’t demonstrate any advantage over Sandy Bridge, and in fact realizes slightly less overall throughput.
During the course of my review, SiSoftware issued Sandra 2012 SP4 based on discussions between Adrian Silasi and I about how the software addresses the clock rates of CPUs with dynamic frequency adjustment like Turbo Boost and Turbo Core.
Because I had all of my data compiled using SP3, though, I decided to use the older build, which reports results based on rated clocks, rather than actual clocks. And really, that decision was made based on the fact that Ivy Bridge and Sandy Bridge demonstrate nearly identical behavior, suggesting very similar cache latencies. Moving forward, I’ll swap over to SP4. But for now, the desired effect is achieved.
