Solved Haswell RAM issue

[Britton30]

The Haswell has a much lower memory read than the Sandy Bridge and I wonder why. Any ideas? The Sandy is OC'ed to 4400 and the Haswell to 4200. The RAM is identical, GSkill Sniper 2133 2x4GB, but is OC'ed to 2200 in the Haswell, same timings, 9-11-10, 1.65V.

 

[kbrady1979]

Quick google search turned this up from Tom's Hardware forums:

Both Haswell processors were tested using 3DMark11, AIDA64, Super Pi, Cinebench 11.5 and Fritz Chess Benchmark programs. The "admin" user also ran tests on Core i7-3770K, and made an attempt to run them on Core i7-4770K, overclocked to 4.5 GHz. Overclocked 4770K, even with water cooling, was not stable enough to complete Cinebench 11.5, and running more CPU intensive benchmarks like Prime95 was out of the question. It seems that, like their Ivy Bridge predecessors, Haswell CPUs exhibit the same problem operating at high or very high temperatures when overclocked. Nevertheless, the user was able to receive some results for Core i7-4770K @ 4.5GHz, and they are presented in the table below:
Benchmarks Core i7-4770K Core i5-4570K Core i7-3770K Core i7-4770K
(overclocked)
3DMark 11 X5085 X4983 X5022 X5013
AIDA64 read 22190 MB/s 20660 MB/s 23247 MB/s 23127 MB/s
AIDA64 write 22673 MB/s 20927 MB/s 20704 MB/s 26125 MB/s
AIDA64 copy 29739 MB/s 28351 MB/s 24907 MB/s 30651 MB/s
Super Pi 9.344 10.187 9.329 8.018
Fritz Chess Benchmark 14398 11476 14683 17461
Cinebench 11.5 CPU 7.82 6.06 7.88
Cinebench 11.5 OpenGL 75.93 68.26 57.74
Looking at the results, one might say that Haswell's CPU performance is pretty much the same as performance of Ivy Bridge counterparts. Haswell chips are slower at reading from memory, but faster at writing to, or copying data in memory. Ivy Bridge-based 3770K was slightly faster in Super Pi, Cinebench 11.5 and Fritz Chess benchmarks. Haswell CPUs do have much higher graphics performance, up to 31% in Cinebench OpenGL test. Lower, than expected, performance of Haswell parts could be attributed to an early motherboard BIOS, and/or a choice of benchmarks. The 4th generation microprocessors should perform better in other tasks, and especially in programs, utilizing new AVX2 and FMA3 instructions.

I would wait for BIOS updates, if there is one out already you might read up on it to see what it changes. Maybe it's something that will be fixed.

Apparently, I didn't scroll down far enough to see this:

The reason it's slower to read from memory is that it's TLB is now more segregated, and even has support for 2GB pages. It's a big change and has a lot more "sorting" to do so to speak.

But that's not to say that some applications can't benefit from this-- actually, these benchmarks show nothing.

Support for 2GB pages is massive. AVX2 support is massive. Transactional memory support is massive. An extra ALU function is massive. Enhanced branch prediction is massive. Significantly more data registers is massive.

Benchmarks don't show all these benefits-- actually, most of them aren't shown. I repeat: Benchmarks cannot show the real perks of Haswell. Benchmarks don't show branch prediction, AVX2, transactional memory, an extra ALU function, and hardly is data register instensive as a lot of them are SIMD arrays and loops like SuperPi.

I don't understand what 3/4 of that stuff means, but I get that there IS a reason why, and that it's not as cool in benchmarks, but benchmarks don't mean much because there are things being done behind the scenes benchmarks can't account for. At least you can feel better that it's not a problem with your hardware.

 

[Britton30]

Thanks for finding that Kelly. Crap I don't understand 7/8 of it, LOL, BUT it does show, as you say, there are benefits that our present benchmarks don't take into account.
Notice that the MaxxMEM doesn't even recognize the RAM with Haswell, while it does on the Sandy. I think Intel has really changed up a lot of "under the table" functions with Haswell we poor minions don't understand.

 

[kbrady1979]

MaxxMEM doesn't like showing data for Ivy Bridge either as we've discussed before:

 

[Britton30]

Ahh, OK, I think I missed that one.

 

[ICIT2LOL]

Gary and Kelly I found this it might not cover the issues you are discussing cos it beyond me apart from i7 and overlocking LOL!!

[Various] Intel Haswell Reviews

 

[Britton30]

Thanks, John, for the info.

 

[Dude]

Gary, like kelly said as Haswell matures you will get BIOS updates to help, but I do think you are getting better performance...not so much in benchmarks, but overall

 

[Britton30]

Oh, I meant to mention there was a BIOS update, two of them in fact, since the board was released. I got the newest one now.

 

[Britton30]

I have found another issue(s) now. memtest 86+ ver. 4.2 reads memory info completely wrong. I don't think it can be trusted to check RAM in a Haswell system. Even in Sandy Bridge memtest can't read the CPU cache info, or not in mine at least.



This is a different monitor and has crappy display.

 

[kbrady1979]

They should update Memtest86+ because it's several generations behind. Also, that's as messed up as a football bat.

 

[Britton30]

True, I think the last update was in 2011.

 

[kbrady1979]

It was 2011. It said it added support for Sandy Bridge, but I think they quit about halfway through that update

 

[Britton30]

What does it show with Ivy? I don't have one of them.

 

[kbrady1979]

Looks like it actually has my information correct in there. Weird...

 

[Britton30]

Most of it, isn't the Ivy a Core Gen 3? It says Core gen 2 in memtest. Hmm, it misses the SPD info too.

 

[kbrady1979]

That's right, I missed that it said Gen 2

 

[ICIT2LOL]

Hmm Gary even I the great plonker can see the nonsense do you think maybe it has something to do with the transfer rates from the Haswell to the RAM??
When I get chance I will run a memtest on the Sandy and Ivy and see what comes up. If your results are anything to go by it's a bit of a worry.

By the by there was some controversy a while ago on a thread I was in with memtest being upgraded from the one we are used to using. I think it was because it was in beta form or something like that.

 

[kbrady1979]

I don't think there is anything to worry about.

 

[Britton30]

John, nothing to fret over, the problem is with memtest, not Haswell. Intel has changed the innards of Ivy and Haswell so that most of the normal benchmarks don't have anything to read from. From what I understand, Haswell handles RAM much better in different ways but isn't reflected in many tests. Memtest only check for errors, it can't tell us how efficient RAM is.