REVIEW: Intel Core i7-3960X and Intel X79 DX79SI

AIDA64 v2.0
We initially ran an earlier version of AIDA64 which gave us some interesting benchmarks to talk about showing that the Core i7-3960X didn't quite perform up to par, that was until an update rolled out and we were notified that there was an AVX optimized version that would improve performance. Upon re-testing everything we saw improvements between 10 and 20% which we would say are pretty significant and as a result we scrapped our benchmarks and reran them with the ones you see below.

AIDA64 CPU Tests

The Core i7-3960X barely edges out the previous-gen i7-990X

Here the Core i7-3960X destroys the competition by a wide margin. 

Once again, the Core i7-3960X wins by a broad margin, almost double the i7-990X

In these four CPU tests that we ran in AIDA we really didn't see anything surprising other than in the Hashing test where the Core i7-3960X didn't actually win by any margin at all, but rather lost to the 12-core Opteron 2431. While the i7-3960X does have 12 threads thanks to its hyper-threading, it appears that the sheer quantity of cores in the Opteron gave it an edge in terms of raw hashing power. However, the workstation version will come with 8 cores enabled and 16 threads, so we'll see how SNB-E will fare against the old Opteron. The new 16-core Opteron debuted today... so AMD still has a chance to remain the winners in this benchmark.

As one can see, in tests like the AES test Intel has made some significant improvements to AES performance in the Core i7-3960X as it doubles the i7-2600K's score and nearly doubles the Core i7 990X's as well. Overall, quite an impressive performance.

AIDA64 FPU Tests

Here we see the 3960X win by about 20% over the 990X

In this test the 3960X beats the 2600K by a whopping 50%

The results continue to be astonishing once again with the i7 3960X beating the i7 2600 by more than 50%

In the four FPU tests, the 3960X once again scores first in almost all of tests by quite a large margin except for one. In the FPU SinJulia test the 3960X actually loses to the older 990X, this could be due in part to the fact that SinJulia is based upon the floating point X87 instruction set rather than the x64 instruction set, so the Core i7-3960X doesn't really improve X87 operation performance. It does, though score about 50% better than the 2600K which could be attributable to the added cores and memory bandwidth.

AIDA64 Memory Tests

Here we can see the $999 3960X just barely beat out the $315 2600K

Interestingly enough, in this test the i7 2600 with only dual-channel 1333 beats out the i7-3690X with quad-channel 1333.

Here we see a veritable tie between the i7-2600K and i7-3960X

In the memory tests, AIDA64 tests for copy, read, write and latency. Surprisingly enough, The Core i7-3960X does not score 1st in the majority of the tests, but rather 2nd in most of them. In the case of the memory copy and memory write tests, it loses to the Core i7-2600K and in the memory latency test it loses to AMD processors. 

The reigning champion is the AMD FX-8150 which comes in at a snappy 45.8ns beating the Athlon64 X2's 47.5ns which also happens to beat the Core i7-3960X. Hopefully Intel can work more on their memory latency with Ivy Bridge as the AMD FX-8150 is currently the memory clock speed champion due to having such a good memory controller (even though it is only dual-channel).

AIDA Cache and Memory Benchmark
Here we take a look at the performance of the Core i7-3960K with its different levels of cache and its memory performance.

You can also see the improvements that we achieved on the memory and cache performance once we overclocked the processor to 4.86GHz. You can notice here that the L1 cache latency actually decreased over Nehalem, i.e. the original Core i7 processor.

As you can see from our overclocks, the Core i7-3960X when overclocked sees huge improvements to L3 cache bandwidth as well as latency in memory and cache. We managed to get sub 1ns latency in the L1 cache with over 151GB/s of bandwidth. This is the first time we're seeing sub 1ns latency after the early days of Pentium 4.