Question about the integrity of i7-3930K and i7-3960X SB-E Benchmarks

[Tyler11440]

It seems odd that they score so much higher than the 990X, yet the real world performance increase is only give or take a few frames per second.

I heard a rumor that PassMark has some sort of maximum real score point, and after it is passed, your score increases nearly exponentially. If this is true, it would certainly explain the SB-E scores.

If not, how are they scoring so high? Is it a bug? Is it real? Thanks in advance!

[Seronx]

Sandy Bridge-E carries the same bug from Sandy Bridge processors

After a certain clock multiplier the clock rate is no longer read this happens in most benchmarks

The lowest scores for i7 3930K-3960X are within the general area of 10K-12K while overclocked results ponder around 14K-20K

One of the biggest reasons the i7 3930K/3960X is better than the i7 990X on Passmark is because of the uops cache
(The uops really only helps with benchmarks that have repetitive instructions)

Other than the Multiplier -> Clock Rate can't be read so it might as well be at stock bug

Intel Core i7-2600 @ 3.40GHz -> 8971 Pmarks
Intel Core 975 @ 3.33GHz -> 7,326 PMarks

8971/4 = 2242.75
7326/4 = 1831.5

uops cache increases throughput in the average of benchmarks by 400 Pmarks

2242.75 x 6 = 13500 Pmarks

Then you have turbo core which places it around 14K but in most real world applications it will be a very small improvement on the i7 990X

I am ignoring Amadhl's Law because I don't know the parallelization factor for PerformanceTest
(20% increase in performance is the best you should actually ever expect when not using AVX)

[David (PassMark)]

We aren't aware of a Sandy Bridge bug (in the CPU itself), but couldn't rule it out either. If you have any details of this it would be interesting. There are currently issues in PerformanceTest release V7.0 (Build 1025) detecting the correct overclocked speeds in some CPUs. We have have several hacks at fixing it over the last year, and have fixed it up for a few CPU models. We were of the opinion that it is technically possible to do it for pretty much all CPUs, but a bit more research was needed. We are looking at it now, but it might be a few more weeks before we have a new release to try.

You can't always compare frames / sec for gaming benchmarks to the PassMark rating. Frames / sec will be dependent on the game, the game settings, the video card & how threaded the game is. It would be a rare game that maxes out a i7-3960X.

There is no such thing as a "maximum real score point" in the software. Nor is there anything exponential in the way the result is calculated. The CPUMark is a weighted average of the 8 CPU tests.

The graph somewhat distorts differences. A 30% increase in performance from 10 to 13 won't be noticed in the graph do to the scale. A 30% increase in performance from 10,000 to 13,000 will look enormous by comparison.

Also the 990X is a bit older so ~90% of the results are 64bit .
The i7-3960X has 100% of it's results being 64bit, and 64bit systems score significantly higher than 32bit.

Looking at the performance in the individual CPU tests, see below, shows that the i7-3960X was a bit better across the board, with the exception of the SSE test. Where the i7-3960X was dramatically better. More than twice as fast in fact.
(Note: The SSE test was also known as Multimedia Instructions test in prior releases. The test is actually a Matrix multiply using SSE instructions using mainly, _mm_mul_ps, _mm_add_ps on single-precision, floating-point.)

We assume this is due to architecture improvements in the i7-3960X. SSE isn't often directly measured in benchmarks, or even used in software all that commonly as it needs to be hand coded in assembler most of the time. So this effect has probably gone somewhat unnoticed (until now).

[David (PassMark)]

Amadhl's Law...

All the CPU tests are 100% in parallel in PT7. There is a new process (and thus thread) started up for each CPU execution unit. In theory there should be linear scaling as you add CPUs / Cores. In practice I suspect cache or RAM bottle necks mean the scaling will taper off as more cores are added.

[Tyler11440]

You can't always compare frames / sec for gaming benchmarks to the PassMark rating. Frames / sec will be dependent on the game, the game settings, the video card & how threaded the game is. It would be a rare game that maxes out a i7-3960X.

Oh yes, I understand. What I meant was, assuming that the same card, game, drivers, and hardware were used, and the only thing swapped out was the CPU, or the closest thing to just a CPU (if the motherboard socket was different). I follow reviewers like Toms Hardware where they do a fine job of changing just one variable at a time to verify validity of concluded results.

There is no such thing as a "maximum real score point" in the software. Nor is there anything exponential in the way the result is calculated. The CPUMark is a weighted average of the 8 CPU tests.

Can a GPU affect the CPU scores in any way? This could partially explain some seemingly undeserved scores. Meaning, people with an expensive high-end Intel i7-...X CPU would be more likely to buy a rockin' card with it before benchmarking versus just the average low-end quad core with integrated graphics.

Also the 990X is a bit older so ~90% of the results are 64bit .
The i7-3960X has 100% of it's results being 64bit, and 64bit systems score significantly higher than 32bit.

So, after new CPUs have had a chance to get to the same full 64bit results, will we see them skipping into the early-mid 10,000's as well?


Also, thank you for taking the time to write such an informative post

[David (PassMark)]

Video card / GPU should have no effect on the CPU score. So this should not impact on the CPU score to any great extent.

The 64bit / 32bit issue is pretty minor here, and would only account for a small part of the difference. But yes, expect average 64bit CPU scores to creep up slowly over time, as we approach 100% 64bit O/S usage.

In this case the major difference is, in my opinion, is that the i7-3960X is simply slightly faster overall, and a whole lot faster when executing SMID/SSE instructions. See this Wikipedia page for background as to what SMID is.
http://en.wikipedia.org/wiki/Streaming_SIMD_Extensions

As the Wikipedia pages says however, "User application uptake of the x86 extensions has been slow with even bare minimum baseline MMX and SSE support (in some cases) not being supported by applications some 10 years after these extensions became commonly available". So only a few applications will benefit from this.

Would be interesting to find another SIMD benchmark and see if the results match ours for these two CPUs. My speculation is that the implementation of the AVX feature in i7-3960X has resulted in better SSE performance. This slide also seems to imply this,
http://www.techpowerup.com/img/11-07-22/110c.jpg
(the i7-3960X is more than twice as fast in some circumstances)

Some interesting details here as well,
http://software.intel.com/en-us/arti...mall-matrices/