Concerning Haswell-E DDR4 Memory Frequency Validation (Why Oscilloscope is Required)

  • News, Articles
  • 17
  • HWBOT

Concerning Haswell-E DDR4 Memory Frequency Validation (Why Oscilloscope is Required)

Author: Pieter-Jan Plaisier

With the Consumer Electronics Show coming up next week, tension in the (corporate) overclocking world is rising. Traditionally we see vendors bring out their A-game closing in on the big public trade-shows ready to claim the best overclocking results. In the past couple of weeks we’ve seen an increased focus on Haswell-E DDR4 memory overclocking. Sadly enough the overclocking fun may be spoiled by a CPU related bug/feature causing CPU-Z validations to be unreliable for memory frequency validation.

With help from our motherboard partners we established Haswell-E to have a similar bug/feature as TL uncovered on the Haswell platform earlier this year.

Memory Strap and Ratio on Haswell-E

There are three parameters affecting the memory frequency on Haswell and Haswell-E platform: the BCLK frequency, the Memory Strap, and the Memory Ratio. Multiplying the three parameters will give you the memory frequency your system is running at. For example, a BCLK frequency of 103 MHz combined with the 133 Strap and 1:8 Ratio will give you a resulting memory frequency of: 103 MHz x 1,33 x 8 = 1096 MHz or DDR4-2192. For Haswell-E there is a wide range of Memory Ratio options going from 1:4 to 1:20, and two Strap options 100 and 133. You can refer to the table below for a full overview of the available Ratios, how they are represented in the CPU-Z application, and they effective DDR4 rated speed using the two Straps.

With the help of our partners, we tested the various combinations of Strap and Ratios. We encountered three different scenarios: 1) The Strap/Ratio combination does not exist or fails to initialize, 2) the Strap/Ratio functions posts and functions as expected, and 3) the Strap/Ratio initializes but fails to complete post most likely due to memory constraint (“53”). As far the testing allows for it, we confirm that Ratios 1:4 to 1:15 work in combination with both 100 and 133 Strap. Ratios 1:17 to 1:20 failed to complete post using Strap 100. Ratios above 1:15 do not exist using Strap 133.

At Run Time Adjustment

Referring back to the original article on Haswell validation problems, the Haswell-E platform exhibits an additional problem. Not only does the Haswell-E platform allow for at run time Memory Strap adjustment, but it also allows for at run time Memory Ratio adjustment! The implications are severe: it means no DDR4 frequency can be validated through CPU-Z’s validator (at this moment).

To understand why, follow the train of thought. Configure the BIOS to boot with settings: BCLK at 100 MHz, Memory Strap at 100, and Memory Ratio of 1:12. This results in an effective operating frequency of 1200 MHz or DDR4-2400. By “flipping the switch” on the register of the Memory Strap, we can change it to 133. The real operating frequency is however NOT adjusted! Software relying on this register will be affected and show a current memory frequency of 1,33x higher. In this case: DDR4-3200. Only by the means of an oscilloscope we can verify the actual memory frequency of DDR4-2400.

This problem was uncovered on the Haswell platform already. On the Haswell-E platform we can also “flip the switch” of the Memory Ratio. To continue using the previous example, we can go up from 1:12 to 1:15. In that case, our frequency validation has jumped from DDR4-2400 to DDR4-4000. In fact, things get worse. It is possible to program the Memory Ratio register to any value up to 1:63 as you can see below.


(click to enlarge – Left: 100 strap with 1:63 Ratio; right: 133 strap with 1:63 Ratio)

Solution

For the moment it does not seem there’s an easy solution for this problem. Software relying on this method to report the memory frequency is affected. At the moment, we don’t know of any software that is accurately reporting the DDR4 memory frequency after adjusting the Strap or Ratio at run time. For memory frequency validation records it seems obvious we must rely on additional verification by means of the oscilloscope, or at live events by verifying correct functioning of the motherboards prior to the record attempts. The staff is currently investigating our options and will update you once we see the full scope of this problem.

For the time being we would recommend to be cautious when seeing any out-of-the-norm DDR4 memory frequency claims by industry(-affiliated) overclockers and/or marketing teams.. Note that even though the methods used for adjusting the registers are common practice in the industry, they are too advanced for common overclockers and enthusiasts to have easy access to. So if your neighbor comes over to brag about his fantastic DDR4 frequency, there’s no need to doubt. Also note that performance figures are not affected by the at run time adjustment, so you can always verify the actual frequency by running a memory performance benchmark.

Finally, I would like to stress the fact that all parties involved are fully cooperative and there’s no reason to doubt any result or achievement published thus far. We are merely reporting on the problem. To be clear: the current First Place for DDR4 Memory Frequency by Toppc of 2016.1 MHz stands as verified using the scope. A higher validation result submitted not so long ago was pulled only because of the verification.

Keep you posted!

The HWBOT Staff.


17

Belgium Massman says:

Well :-/

South Africa Vivi says:

Dont worry! Its the right way

Canada Trouffman says:

just more difficult to validate :) Glad that everyone is playng by the rules :) !

Canada Joa3d43 says:

...that's ok, I just run my DDR4 on Win 8 / 8.1 w/ that RTC issue > the bugs might cancel each other out ;-) Happy New Year

United States Splave says:

Maybe remove points for it. I can't afford an oscilloscope

K404 says:

Splave said: Maybe remove points for it.

I can't afford an oscilloscope


I was trying to think of a reply for the thread, but Allen's is just fine.

If it needs expensive (for quality) additional hardware for verification, it's a broken benchmark.

Will the rules state what magnification is permitted? How will that be enforced? ;)

Germany der8auer says:

I have to agree with Kenny and Allen. I do own an oscilloscope but we can't expect people to buy them just for a memory validation.

Norway knopflerbruce says:

It's no disaster if certain platforms are not accepted without oscilloscope verification. Sooner or later there will be other DDR4 setups that can be used - maybe AMD APUs, or whatever.

Belgium Massman says:

Only oscilloscope for "engineering records" is not really an option either. Our only hope for now is that CPU-Z can come up with something around this issue. Removing points for DDR4 memory clock is a good idea though.

zeropluszero says:

Joa3d43 said: ...that's ok, I just run my DDR4 on Win 8 / 8.1 w/ that RTC issue > the bugs might cancel each other out ;-) Happy New Year


LOL'd IRL.

Hiwa says:

Is there any cheap oscilloscope can do the same?
Found some on eBay for 50/60 euros.

Germany der8auer says:

You can get oscilloscopes which are quite okay for around 200-300€. Would propbably better if you straight invest in a better one.

Indonesia Lucky_n00b says:

Well, if the scopes are now a requirement for record, and there's affordable one, I'm interested in learning the proper method of measuring the DDR Mhz using oscilloscopes :)

TaPaKaH says:

Another $1k investment necessary? Talk about this hobby being cheap... :D

K404 says:

I don't think that buying an oscilloscope is a "legitimate" workaround for this problem and if people start buying them, then we have another financial barrier built in to the hobby. Massman recently posted asking "How to engage the Extreme Rookie Community?" I imagine a good answer is.... not setting a precedent of needing to buy an oscilloscope to compete

Czech Republic RRainbo says:

No guys :D No chance to probe 2GHz differential signal cheaply. You will not see it on low bandwith scopes. You need very fast scope for it. Like 5 times more bandwith to see the signal properly. To just measure frequency you will still need bandwith to be more than your measured signal so something like 4GHz scope to have reserve when DDR4 will mature.

And matching differential probes are not cheap either. So for secondhand from ebay expect more then $5k+ scope and $1k+ probe. Hicookies measure setup contains TDS 7704B + P7520 thats together $40k atleast. Suddenly 5960x looks cheap :D

And how is it probed? There is small "connector" soldered on test points look on pics

http://electronicdesign.com/site-files/electronicdesign.com/files/archive/electronicdesign.com/content/content/62344/62344-fig-2.jpg
http://dev.xdevs.com/attachments/download/626/dqs1.jpg

Brazil Rbuass says:

Maybe soon you will need a tutorial for osciloscope validations... hehehe

Please log in or register to comment.

Leave a Reply: (BBCODE allowed: [B], [QUOTE], [I], [URL], [IMG],...)