Significantly Higher Memory Mark Score Going from 4670K to 8350?

[Espionage724]

I had an Intel i5-4670K and ran my memory at stock values (9-9-9-21 @ 1333MHz). I went over to an AMD FX-8350 with the same values, and PassMark showed a far lower score.

As for the numbers:

4670K: 2400
FX-8350: 1356
Phenom II X4: 1132

I'm not entirely certain on other values aside from the timing, speed, and command rate; but those three values should have been the same across all motherboard/CPUs (I didn't go out of my way to adjust them basically).

I'm wondering if PassMark may have made a mistake with the 4670K memory test, or perhaps if there's some kind of biased testing occurring, or maybe the 4670K just has a significantly better way of handling memory?

[David (PassMark)]

The newer Intel chips have a better memory controller.

Basically the AMD chips can't fully use the available bandwidth of high end DDR3 ram. (and can't use DDR4 at all at the moment). The exception to this is memory usage from integrated GPUs. Which can make use of faster RAM.

[claurie]

I had an Intel i5-4670K and ran my memory at stock values (9-9-9-21 @ 1333MHz). I went over to an AMD FX-8350 with the same values, and PassMark showed a far lower score.

As for the numbers:

4670K: 2400
FX-8350: 1356
Phenom II X4: 1132

I'm not entirely certain on other values aside from the timing, speed, and command rate; but those three values should have been the same across all motherboard/CPUs (I didn't go out of my way to adjust them basically).

I'm wondering if PassMark may have made a mistake with the 4670K memory test, or perhaps if there's some kind of biased testing occurring, or maybe the 4670K just has a significantly better way of handling memory?

Just take a look at the Windows integrated memory benchmark WinSAT and you will see a much lower difference (7.8 vs 7.6 in dual chanel mode : all your CPUs have 2 memory channels).

Such difference can only be benchmark software related (you can call it biased but if it was not intentional it can also be exponential scoring or any other thing like running in the cache instead of the memory in the memory test, or dividing the score per the number of cores (the more core you have, the same theoretical bandwidth you share with all of them)).

Max theoretical bandwidth for your memory is easy to find, and integrated memory controllers have different efficiencies that can also depend on core architecture, but cannot be 40% : that's just not serious.

[David (PassMark)]

different efficiencies that can also depend on core architecture, but cannot be 40%

It is really that hard to believe that two different CPUs have a 40% speed difference.

What is true however is that this doesn't often translate into a difference in application performance. Most applications are doing a mix of activities on the CPU, disk & RAM. So the performance of the memory controller doesn't have a huge impact for many applications (it manages to keep up better as the load isn't as intense as in the benchmark).