Author Topic: DDR4 DIMM Lies  (Read 10386 times)

0 Members and 1 Guest are viewing this topic.

Offline gnifTopic starter

  • Administrator
  • *****
  • Posts: 1672
  • Country: au
DDR4 DIMM Lies
« on: March 22, 2017, 02:43:47 am »
Hi All,

I just upgraded to the Ryzen 1700X, with which I bought 16GB (4x4GB) DDR4 dimms rated at DDR-2400. On first boot the system wouldn't run at 2400, but instead rebooted several times until it finally came up (obviously detecting the ram) at DDR4-1833, forcing the RAM to run at DDR4-2400 caused the system to become unstable, even when manually setting the CAS/RAS/Precharge, etc... to their maximum values. The ram runs perfectly fine at DDR4-2133 however, but this is not what I bought nor what is plastered all over the ram, and the box.

So, I looked up the datasheet for the RAM IC itself: HERE
The part is marked:  H5AN4G6NMFR-TFC
The datasheet clearly states that parts graded with TF are only 2133 MHz parts.

Later I found out that these modules are JEDEC-2133 but XMP-2400 (Intel's Extreme Memory Profile), which is why they are selling them as DDR4-2400, this seems like it is common across the ram today. Anyway, so I figured I would get some faster RAM, knowing that it could have XMP profiles on the RAM and that is what it is being sold with.. so I purchased 16GBx2 DDR4-3200. I have not yet even installed it and figured I would save myself the trouble and inspect the ICs under the headspreader (which I am now convinced is not for heat, but to hide the ICs part).

The parts are: H5AN8G8NMFR-TFC
Obviously it's already suspicious.
HERE is its datasheet (which BTW hynx have removed from their website, google still has a cache of the URL).
Again, TF are 2133 MHz parts.

1) Am I missing something here?, or
2) Is this a major conspiracy in that we are all being sold overclocked RAM from factory?

I am going to now install this ram and see what happens, but it seems that manufacturers are lying about the performance of the ram by not stating its true rated speed, but it's overclocked (XMP) speed.

For the record the two different packs of ram are:

4x4GB = AVEXIR Platinum Series (AVD4u24001604G-4M)
2x8GB = G.Skill Ripjaws V (F4-3200C16D-16GVKB)
« Last Edit: March 22, 2017, 02:48:51 am by gnif »
 
The following users thanked this post: bitwelder

Offline gnifTopic starter

  • Administrator
  • *****
  • Posts: 1672
  • Country: au
Re: DDR4 DIMM Lies
« Reply #1 on: March 22, 2017, 03:00:47 am »
I honestly don't know at this point as I have not had the system long enough (literally only a few days). I do know that between 1833MHz and 2133MHz the difference is enormous in games, I see huge latency spikes on the lower frequency.
 

Offline mariush

  • Super Contributor
  • ***
  • Posts: 4983
  • Country: ro
  • .
Re: DDR4 DIMM Lies
« Reply #2 on: March 22, 2017, 03:06:10 am »
Ryzen really likes fast ram , as the infinity fabric inside the processor also runs at the speed of the ddr4 controller from what i heard (but it's 256 bit wide, 20 GB/s fabric or something like that ). So if your memory sticks run at 2400 Mhz, the fabric inside runs at 1200 mhz ( 2x1200 mhz= 2400 mhz)

Infinity fabric connects the two core complexes ( each core complex has 4 cores and 8 threads and access to the 16 MB of level 3 cache) and the pci-e complex (all the links) which includes the connection to the chipset (connection to chipset is a pci-e v3.0 x4 link).

officially ryzen supports 2667 Mhz with 2 single sided memory sticks (all 4 GB sticks and some 8GB sticks are SS) , one per channel , 2400 Mhz with 2 double sided (pretty much all 8 GB sticks and higher) memory sticks, one per channel  and then the maximum drops to 2133 Mhz / 1866 Mhz when you have 4 sticks installed.

Unofficially, they're going to support over 3200 Mhz, some got them working with up to 3500 Mhz ...  but the launch of the processors was rushed and the BIOSes and the code inside the processors had some memory settings locked to some standard values which made the memory compatibility suck at lunch.

It's fixable through bios updates for various motherboards and through microcode uploads (AGESA code) in the processors (which is also done by flashing the bios) ...

So don't be very surprised if you can't get your memory sticks to work at some high frequencies (for ex your 3200 mhz may only work at 2933 mhz and higher timings for now), it may not be the memory sticks' fault, it could be a side effect of the cpu/motherboard memory support that will be fixed through bios updates.  Some motherboards couldn't even run sticks at 2400 Mhz, which looks like it could be your case as well.
 
As for that TC thing it could be that the ram maker binned the chips more thoroughly than hynix / tsmc / samsung who maybe simply stopped testing them when they determined they worked at 1200 Mhz or some value.  Or maybe hynix binned them using some lower latency parameters and found the chips not being stable at that high frequency , but the ram maker increased the parameters and found them stable at higher frequencies.


// sorry for spelling mistakes / lack of upper and lowercase, using a temporary keyboard which has very mushy and small keys, poured some Pepsi over my default one

ps2

If you're using Windows 10 the gaming experience may suck more than with other OSes because MS still has to patch the scheduler to see the cores and threads properly and be aware of the two core complexes.
Basically the two core complexes are sort of like (but not really) those Intel processors that had two cpu dies connected together on a substrate under the metal lid.  Each has 4 cores and 8 threads and each core has its own L1 cache and 512K of L2 cache and access to the 16 MB L3 cache. 
But the problem is that an applications that uses lots of threads could have one thread talk to a thread that's on the other core complex, in which case there's a latency involved (moving data from one core complex to the other through the slower L3 cache) ... basically it's something like instead of 0.5 ns when two threads talk inside one core complex, it's about 8-10 ns to talk between threads on two core complexes.
The current sheduler in Windows 10 is not smart enough to set threads on real cores in the first place, then move to the virtual threads (the hyperthreaded ones), the scheduler detects now as 16 cores if i'm not mistaken... and it just puts the threads on random cores/threads, without trying to group as best as possible threads of one application to one core complex. 
Generally, applications that use lots of threads like x264 for example or image processing apps or things massively parallel load chunks of something in each thread and give threads some work, so theres little chat (if any) between threads.
But with games,  it's more difficult as some thread often has to poll other threads for updates (like physics thread waiting for AI, or some other thread waiting for directx thread etc )
« Last Edit: March 22, 2017, 03:16:48 am by mariush »
 

Offline senso

  • Frequent Contributor
  • **
  • Posts: 951
  • Country: pt
    • My AVR tutorials
Re: DDR4 DIMM Lies
« Reply #3 on: March 22, 2017, 03:13:17 am »
"Easy" fix, swap for a 2x8GB kit and it will work, Ryzen for now as problems with high clock RAM's over 2 sticks, and the controller is dual channel, so no gain in going to 4 sticks, that or wait for BIOS updates from the mobo manufacturer, or go over win-raid and see if anyone already as hacked some modded BIOS already.
 

Offline BradC

  • Super Contributor
  • ***
  • Posts: 2104
  • Country: au
Re: DDR4 DIMM Lies
« Reply #4 on: March 22, 2017, 03:13:26 am »
Keep us updated. I've got a backorder on a new 1800X with GSkill Ripjaws (4x8GB). I picked part numbers that were specifically listed on the ASUS Ram QVL so they *should* work at rated speed (2400).
 

Offline gnifTopic starter

  • Administrator
  • *****
  • Posts: 1672
  • Country: au
Re: DDR4 DIMM Lies
« Reply #5 on: March 22, 2017, 03:23:32 am »
My main issue with this isn't what the Ryzen will run at, it is that they (the DIMM manufacturer) have lied about it's performance. For example, if you opt to buy 3200 ram and want to overclock, etc... you have no headroom, it is already overclocked. Or worse, you buy it where stability is the requirement, and the ram is overclocked... would you trust it?

I know Ryzen has issues with RAM at current, and you are correct that it doesn't seem to like 4x as if I pull two out the system then clocks at 2400, but the system fails to post 10% of the time. But this is not the issue here, the issue is the marketing BS.
« Last Edit: March 22, 2017, 03:25:43 am by gnif »
 

Offline Monkeh

  • Super Contributor
  • ***
  • Posts: 7990
  • Country: gb
Re: DDR4 DIMM Lies
« Reply #6 on: March 22, 2017, 03:26:01 am »
1) Am I missing something here?

Nope.

Quote
2) Is this a major conspiracy in that we are all being sold overclocked RAM from factory?

When you buy the fancy brand stuff with shiny heatspreaders and XMP profiles - yes.

Go buy some DDR4-2400 from Crucial. It will be proper. Unlike the popular gamer brands, Crucial effectively makes RAM (being Micron, one of the largest manufacturers thereof in the world).

Whether that'll work reliably with a first-gen Ryzen system is another matter.
 
The following users thanked this post: orolo

Offline gnifTopic starter

  • Administrator
  • *****
  • Posts: 1672
  • Country: au
Re: DDR4 DIMM Lies
« Reply #7 on: March 22, 2017, 03:27:16 am »
If you're using Windows 10 the gaming experience may suck more than with other OSes because MS still has to patch the scheduler to see the cores and threads properly and be aware of the two core complexes.
Basically the two core complexes are sort of like (but not really) those Intel processors that had two cpu dies connected together on a substrate under the metal lid.  Each has 4 cores and 8 threads and each core has its own L1 cache and 512K of L2 cache and access to the 16 MB L3 cache. 
But the problem is that an applications that uses lots of threads could have one thread talk to a thread that's on the other core complex, in which case there's a latency involved (moving data from one core complex to the other through the slower L3 cache) ... basically it's something like instead of 0.5 ns when two threads talk inside one core complex, it's about 8-10 ns to talk between threads on two core complexes.
The current sheduler in Windows 10 is not smart enough to set threads on real cores in the first place, then move to the virtual threads (the hyperthreaded ones), the scheduler detects now as 16 cores if i'm not mistaken... and it just puts the threads on random cores/threads, without trying to group as best as possible threads of one application to one core complex. 
Generally, applications that use lots of threads like x264 for example or image processing apps or things massively parallel load chunks of something in each thread and give threads some work, so theres little chat (if any) between threads.
But with games,  it's more difficult as some thread often has to poll other threads for updates (like physics thread waiting for AI, or some other thread waiting for directx thread etc )

Please see: https://community.amd.com/community/gaming/blog/2017/03/13/amd-ryzen-community-update?sf62107357=1
Those reports were based on old software that didn't know how to interrogate windows 10 for the information correctly.
 

Offline gnifTopic starter

  • Administrator
  • *****
  • Posts: 1672
  • Country: au
Re: DDR4 DIMM Lies
« Reply #8 on: March 22, 2017, 03:38:28 am »
Ryzen is infamous for its RAM compatibility.
It's mostly criticized for its bad performance (due to Windows kernel task scheduler bug, AMD has fixed this already), very limited over-clock-ability and RAM compatibility.
If you want to run it at high RAM clock, buy Gskill Flare-X RAM, they are specially designed and binned to match Ryzen's poorly designed RAM controller's timings.
You can get up to 64GB of RAM at 3200MT/s, and 32GB at 3466MT/s, both require 4 channels to be used (which means, with currently available best timing RAM, you only get 8GB/stick at 3466MT/s or 16GB/stick at 3200MT/s).
There are rumors saying AMD will tweak BIOS to support wider range of RAMs, but I think it is a silicon issue, and for the current silicon revision, this is the best you can have.
You won't get 128GB at 3000MT/s or even 3200MT/s like with i7x or E5-16xx.

Again, this is not about AMD/Ryzen, this is about the marketing BS stating XMP speeds rather then the true JEDEC speeds on the ram. And after seeing that G.Skill are doing this here there is no way I would trust them with more of my money. I want to know the true JEDEC speed, not the overclocked XMP junk.
 

Offline gnifTopic starter

  • Administrator
  • *****
  • Posts: 1672
  • Country: au
Re: DDR4 DIMM Lies
« Reply #9 on: March 22, 2017, 03:52:04 am »
Surprise surprise... the 3200 wont post at 3200... only 2133. There is an option to enable the XMP profile in the BIOS now, which when done, the system wont post. Interestingly they also advertise it is 20% lower voltage at 1.20V, they all seem to, but the DDR-3200 XMP profile raises the voltage to 1.350V, so another lie.
 

Offline Monkeh

  • Super Contributor
  • ***
  • Posts: 7990
  • Country: gb
Re: DDR4 DIMM Lies
« Reply #10 on: March 22, 2017, 03:53:48 am »
Surprise surprise... the 3200 wont post at 3200... only 2133. There is an option to enable the XMP profile in the BIOS now, which when done, the system wont post.

This is more likely to be the rest of the system at fault. The RAM will have been tested (by whoever packaged it) to function at the higher speeds. It's still a crock, but at least it should be a functional crock.

Quote
Interestingly they also advertise it is 20% lower voltage at 1.20V, they all seem to, but the DDR-3200 XMP profile raises the voltage to 1.350V, so another lie.

Well see, that's the thing, if you want RAM which didn't work at 300MHz to run at 300MHz, you're going to need to tweak the voltage a bit. Remember before XMP profiles, you'd get the timings and raised voltages on the sticker..
« Last Edit: March 22, 2017, 03:57:16 am by Monkeh »
 

Offline gnifTopic starter

  • Administrator
  • *****
  • Posts: 1672
  • Country: au
Re: DDR4 DIMM Lies
« Reply #11 on: March 22, 2017, 04:27:43 am »
Surprise surprise... the 3200 wont post at 3200... only 2133. There is an option to enable the XMP profile in the BIOS now, which when done, the system wont post.

This is more likely to be the rest of the system at fault. The RAM will have been tested (by whoever packaged it) to function at the higher speeds. It's still a crock, but at least it should be a functional crock.

So its my systems fault that the RAM is falsely advertised?... no where on the DDR module, or packet does it state it is an XMP profile, nor does it state its actual speed...

Quote
Interestingly they also advertise it is 20% lower voltage at 1.20V, they all seem to, but the DDR-3200 XMP profile raises the voltage to 1.350V, so another lie.

Well see, that's the thing, if you want RAM which didn't work at 300MHz to run at 300MHz, you're going to need to tweak the voltage a bit. Remember before XMP profiles, you'd get the timings and raised voltages on the sticker..

I completely understand, the voltage needs to be bumped to overclock in some cases to account for leakage, etc.... but they state it's 20% lower then DDR3 voltage, of 1.5V, 1.35 is not 20% lower, its misleading.

I am only able to achieve a post at 2400MHz with XMP enabled, anything faster and the system wont post.

Here is what we know for certain:

  • The RAM modules are really only binned at 2133 by Hynx
  • Hynx also grade their memory for 2666MHz and mark them appropriately according to the datasheet
  • The packaging of the DDR does not state the speed is an overclocked/XMP profile
  • These ram modules work on Intel systems... but not on AMDs flagship.

Here is what we know of intel:

  • They produce the 'Best' compiler for C/C++, but it cripples performance for non Intel platforms
  • They paid HP/Dell/Compaq/Toshiba under the table to prevent AMD processors being used in new builds
  • They lied about the P4 performance and even paid off some benchmark companies to lie about the P4 performance figures

Why people trust them IMO is beyond me. It is not a large leap to think that perhaps Intel are doing something dodgy with their XMP that favors Intel, some difference in timing, clock control, etc... or even outright lying about reported memory clock speeds.

It is mighty strange that AMD are having issues with DDR4 on their flagship CPU, which their entire company is relying on to rescue them.
 

Offline Monkeh

  • Super Contributor
  • ***
  • Posts: 7990
  • Country: gb
Re: DDR4 DIMM Lies
« Reply #12 on: March 22, 2017, 04:33:24 am »
Surprise surprise... the 3200 wont post at 3200... only 2133. There is an option to enable the XMP profile in the BIOS now, which when done, the system wont post.

This is more likely to be the rest of the system at fault. The RAM will have been tested (by whoever packaged it) to function at the higher speeds. It's still a crock, but at least it should be a functional crock.

So its my systems fault that the RAM is falsely advertised?... no where on the DDR module, or packet does it state it is an XMP profile, nor does it state its actual speed...

No, that it doesn't boot. As I said, it's a crock, but it's a crock which should work.

Quote
Quote
Interestingly they also advertise it is 20% lower voltage at 1.20V, they all seem to, but the DDR-3200 XMP profile raises the voltage to 1.350V, so another lie.

Well see, that's the thing, if you want RAM which didn't work at 300MHz to run at 300MHz, you're going to need to tweak the voltage a bit. Remember before XMP profiles, you'd get the timings and raised voltages on the sticker..

I completely understand, the voltage needs to be bumped to overclock in some cases to account for leakage, etc.... but they state it's 20% lower then DDR3 voltage, of 1.5V, 1.35 is not 20% lower, its misleading.

Yes, it's bullshit marketing. Again, this is what you get with gamer crap - go buy stuff which doesn't pretend it needs a heatsink!

Quote
Here is what we know for certain:

  • The RAM modules are really only binned at 2133 by Hynx
  • Hynx also grade their memory for 2666MHz and mark them appropriately according to the datasheet
  • The packaging of the DDR does not state the speed is an overclocked/XMP profile
  • These ram modules work on Intel systems... but not on AMDs flagship.

Here is what we know of intel:

  • They produce the 'Best' compiler for C/C++, but it cripples performance for non Intel platforms
  • They paid HP/Dell/Compaq/Toshiba under the table to prevent AMD processors being used in new builds
  • They lied about the P4 performance and even paid off some benchmark companies to lie about the P4 performance figures

Why people trust them IMO is beyond me. It is not a large leap to think that perhaps Intel are doing something dodgy with their XMP that favors Intel, some difference in timing, clock control, etc... or even outright lying about reported memory clock speeds.

It is mighty strange that AMD are having issues with DDR4 on their flagship CPU, which their entire company is relying on to rescue them.

So it's an Intel conspiracy?

You have a brand new board, with a brand new CPU, and barely-tested-in-the-wild firmware. Problems are not unexpected. Intel CPUs have had similar issues previously. The fact that it won't boot any RAM at high clocks suggests a system issue. XMP is nothing but another set of timings in the EEPROM - all the work is done by the memory controller.

If you're not careful with the conspiracy theories we'll end up with TerraHertz gracing this thread, and then all attempts at rational discussion will be derailed.
« Last Edit: March 22, 2017, 04:40:41 am by Monkeh »
 

Offline gnifTopic starter

  • Administrator
  • *****
  • Posts: 1672
  • Country: au
Re: DDR4 DIMM Lies
« Reply #13 on: March 22, 2017, 05:19:26 am »
So it's an Intel conspiracy?

You have a brand new board, with a brand new CPU, and barely-tested-in-the-wild firmware. Problems are not unexpected. Intel CPUs have had similar issues previously. The fact that it won't boot any RAM at high clocks suggests a system issue. XMP is nothing but another set of timings in the EEPROM - all the work is done by the memory controller.

If you're not careful with the conspiracy theories we'll end up with TerraHertz gracing this thread, and then all attempts at rational discussion will be derailed.

Yes, I realize this also, but I cant help but think it's possible. I have been messing around with timings, if I use the XMP profile and override the clock, I can post at 2400MHz fine, if I set the timings and voltage to the same as the XMP profile, but leave XMP off, it wont post... there is clearly something more to XMP then just the timings. CPU-Z reports Bank Cycle Time (tRC) which there are no setting for, with XMP enabled this is set to 57, without it is 51.

With XMP enabled @ 2400MHz I have managed to go from 18-18-18-38 to 12-12-12-27. This seems like a pretty huge drop in latency, makes me wonder if intel are calculating the throughput and using that to calculate a frequency. IE: DDR = base clock * 2, which is also a crock.
 

Offline forrestc

  • Supporter
  • ****
  • Posts: 646
  • Country: us
Re: DDR4 DIMM Lies
« Reply #14 on: March 22, 2017, 05:37:56 am »
2) Is this a major conspiracy in that we are all being sold overclocked RAM from factory?

There is a reason why I, as a general rule, only buy memory from Crucial (Micron) and specifically the type of memory recommended by their memory selector.  Back when I was building PC's for part of my living, I for quite a while kept looking for 'bargains' in the memory department.  At some point I realized it just wasn't worth the pain and suffering that questionable memory causes, and decided to switch to an 'only crucial' policy.   Sure has saved me a lot of pain over the years - I don't remember the last time I had a memory-related problem.

 

Offline Monkeh

  • Super Contributor
  • ***
  • Posts: 7990
  • Country: gb
Re: DDR4 DIMM Lies
« Reply #15 on: March 22, 2017, 05:42:04 am »
So it's an Intel conspiracy?

You have a brand new board, with a brand new CPU, and barely-tested-in-the-wild firmware. Problems are not unexpected. Intel CPUs have had similar issues previously. The fact that it won't boot any RAM at high clocks suggests a system issue. XMP is nothing but another set of timings in the EEPROM - all the work is done by the memory controller.

If you're not careful with the conspiracy theories we'll end up with TerraHertz gracing this thread, and then all attempts at rational discussion will be derailed.

Yes, I realize this also, but I cant help but think it's possible. I have been messing around with timings, if I use the XMP profile and override the clock, I can post at 2400MHz fine, if I set the timings and voltage to the same as the XMP profile, but leave XMP off, it wont post... there is clearly something more to XMP then just the timings. CPU-Z reports Bank Cycle Time (tRC) which there are no setting for, with XMP enabled this is set to 57, without it is 51.

With XMP enabled @ 2400MHz I have managed to go from 18-18-18-38 to 12-12-12-27. This seems like a pretty huge drop in latency, makes me wonder if intel are calculating the throughput and using that to calculate a frequency. IE: DDR = base clock * 2, which is also a crock.

XMP really is nothing but timings and voltages. It's all driven by the memory controller. That does not mean everything it sets is exposed in $INSERT_CRAPPY_FIRMWARE_HERE

I'm pretty sure they're not lying about the speed (ps. MT/s, not MHz, there's a distinct difference). You'll just need to wait for AMD and Asus/Gigabyte/whatever equally terrible brand to get their crap together and get the controller working at higher clocks - or they'll admit defeat and you'll have to buy into the next silicon revision..
 

Offline gnifTopic starter

  • Administrator
  • *****
  • Posts: 1672
  • Country: au
Re: DDR4 DIMM Lies
« Reply #16 on: March 22, 2017, 05:46:04 am »
XMP really is nothing but timings and voltages. It's all driven by the memory controller. That does not mean everything it sets is exposed in $INSERT_CRAPPY_FIRMWARE_HERE

I'm pretty sure they're not lying about the speed (ps. MT/s, not MHz, there's a distinct difference). You'll just need to wait for AMD and Asus/Gigabyte/whatever equally terrible brand to get their crap together and get the controller working at higher clocks - or they'll admit defeat and you'll have to buy into the next silicon revision..

You are likely correct, I am just very disappointed to see how much "gamer" junk is now on the market. Enabling XMP is clearly setting something else unknown, as I stated the board wont post at the XMP timings (the ones I can see and set) without XMP enabled.

I guess it is a wait and see, but lesson learned, watch out for memory that markets it's XMP timings only (at this point anything > 2400 seems to be it).

For anyone interested I figure I will post my performance stats:

On stock voltage the 1700X will clock at 3.8GHz stable, it will run at 3.9GHz until loaded.
I am running water cooling, no changes to the setup since my FX-8350, same water block, etc.
If I push the voltage to 1.45V I can clock at 4GHz stable, but I don't like the voltage bump just for that bit of performance, it's already fast enough.
This is on an ASrock AB350 Pro4, not exactly a high end board, but seems to run very well.
CPU-Z
« Last Edit: March 22, 2017, 06:02:01 am by gnif »
 

Offline Monkeh

  • Super Contributor
  • ***
  • Posts: 7990
  • Country: gb
Re: DDR4 DIMM Lies
« Reply #17 on: March 22, 2017, 05:47:50 am »
As forrestc also mentioned: Stick to Crucial. Just the plain stuff, not the Ballistix. Just plain old naked Crucial. Yeah, it's not always the cheapest, the range is somewhat limited, but it's just RAM. And it just works. .. and when it doesn't it's got a lifetime warranty and they're very prompt about it.

It usually doesn't seem to be on the supported lists with consumer boards - but those lists only exist because of stupid XMP profiles and broken boards.
« Last Edit: March 22, 2017, 05:49:53 am by Monkeh »
 

Offline Jeroen3

  • Super Contributor
  • ***
  • Posts: 4067
  • Country: nl
  • Embedded Engineer
    • jeroen3.nl
Re: DDR4 DIMM Lies
« Reply #18 on: March 22, 2017, 06:49:17 am »
Buy sticks without heatspreader.

They might not be the cheapest around. But one bad bit is enough to spray you with bluescreens.
 

Offline Mr. Scram

  • Super Contributor
  • ***
  • Posts: 9810
  • Country: 00
  • Display aficionado
Re: DDR4 DIMM Lies
« Reply #19 on: March 22, 2017, 08:07:57 pm »
Again, this is not about AMD/Ryzen, this is about the marketing BS stating XMP speeds rather then the true JEDEC speeds on the ram. And after seeing that G.Skill are doing this here there is no way I would trust them with more of my money. I want to know the true JEDEC speed, not the overclocked XMP junk.
Did you consider that the manufacturer might have tested and binned the memory? Just like chip manufacturers test chips and bin them accordingly, they could very well test the chips and sell those that work at higher speed at those higher speeds. Dave recently did a video on running chips outside their manufacturers specifications. The rating is a minimum and often leaves some room in part of the production run.

The thing probably is that Ryzen is rather new. These chips were tested and work with Intel CPUs at those speeds. Unfortunately, Ryzen does not play well with XMP. Considering how dependent Ryzen is on memory speeds, it is likely that sets of specifically Ryzen compatible memory will soon be brought to market. If only for the marketing wank advantage.
 

Offline Monkeh

  • Super Contributor
  • ***
  • Posts: 7990
  • Country: gb
Re: DDR4 DIMM Lies
« Reply #20 on: March 22, 2017, 08:17:28 pm »
Again, this is not about AMD/Ryzen, this is about the marketing BS stating XMP speeds rather then the true JEDEC speeds on the ram. And after seeing that G.Skill are doing this here there is no way I would trust them with more of my money. I want to know the true JEDEC speed, not the overclocked XMP junk.
Did you consider that the manufacturer might have tested and binned the memory? Just like chip manufacturers test chips and bin them accordingly, they could very well test the chips and sell those that work at higher speed at those higher speeds. Dave recently did a video on running chips outside their manufacturers specifications. The rating is a minimum and often leaves some room in part of the production run.

Well of course they bin them - the point is that they're upselling a product they don't even manufacture. And they regularly do it by cranking the voltage past normal levels.
 
The following users thanked this post: gnif, thm_w

Offline Jeroen3

  • Super Contributor
  • ***
  • Posts: 4067
  • Country: nl
  • Embedded Engineer
    • jeroen3.nl
Re: DDR4 DIMM Lies
« Reply #21 on: March 23, 2017, 07:56:05 am »
Did you consider that the manufacturer might have tested and binned the memory?
Might? You mean do. They've made special machines to classify the products.

 
The following users thanked this post: Mr. Scram

Offline technogeeky

  • Frequent Contributor
  • **
  • Posts: 555
  • Country: us
  • Older New "New Player" Player Playa'
Re: DDR4 DIMM Lies
« Reply #22 on: March 23, 2017, 09:07:37 am »
I haven't built a new PC in a while, and I assume OP has built computers more recently than I have. But I think this might be a cause of safe defaulting with SPD versus the "real" settings in whatever extended memory profile you are talking about.

But perhaps people are just selling RAM that has to have a very specific combination of settings to work at some particular (faster) frequency? I would guess it all starts with the higher voltage. And then you'll bring on the more and more relaxed timings.

Pick some very slow memory speed, boot the computer, and then dump all of timing settings profiles (both regular and extended, evidently).

I know the last time I built a computer, the SPD settings profile for the memory were vastly lower than what the memory was sold as. In my case, there weren't any extended SPD settings, I had to enter some exact settings as listed on Crucial's forum. But when I did, the system was stable at the high freq.

 

Offline BradC

  • Super Contributor
  • ***
  • Posts: 2104
  • Country: au
Re: DDR4 DIMM Lies
« Reply #23 on: March 23, 2017, 09:27:40 am »
I haven't built a new PC in a while, and I assume OP has built computers more recently than I have. But I think this might be a cause of safe defaulting with SPD versus the "real" settings in whatever extended memory profile you are talking about.

From my limited reading it would appear that most commonly available DDR4 chips are rated at about 2100 or 2400. The "art" as it were is in testing, tuning and binning the chips to assemble factory overclocked modules to flog to the general public. So yeah, pretty much anything over those values and you are buying something that has been tested to work at the sticker frequency, but came off the wafer with much more moderate targets.
 

Offline madires

  • Super Contributor
  • ***
  • Posts: 7695
  • Country: de
  • A qualified hobbyist ;)
Re: DDR4 DIMM Lies
« Reply #24 on: March 23, 2017, 11:26:35 am »
This is on an ASrock AB350 Pro4, not exactly a high end board, but seems to run very well.

I don't know if ASrock has become better the last few years, but they've built cheap mainboards, cheap in all regards.
 

Offline mrflibble

  • Super Contributor
  • ***
  • Posts: 2051
  • Country: nl
Re: DDR4 DIMM Lies
« Reply #25 on: March 23, 2017, 11:31:34 am »

So it's an Intel conspiracy?
Never suspect malice when incompetence will suffice.
Never suspect malice when incompetence or marketing bullshit will suffice.

Quote
You have a brand new board, with a brand new CPU, and barely-tested-in-the-wild firmware. Problems are not unexpected.
Pretty much this...

Don't get me wrong, when the jedec profiles are lacking, having an XMP profile is convenient. You can consider it a handy lookup table and you don't have to enter all them numbers into the bios settings like you had to in the dark ages. But ... if you have to use an XMP profile because the jedec profiles are not there yet, that's usually a hint that you are running those memory chippies on settings the fab would probably frown upon.

You had a similar situation years ago with DDR3. You couldn't get nicely priced fast modules that not only were fast & cheap, but also ran on 1.50 V as intended. You could however get fast & cheap if you accepted that the memory stick vendor already did the overvolting to 1.65 V for you. A couple of years later 1.50 V modules were plentiful. And ditto for timing profiles vs price+availability.

So looks like we're currently still in the era where it makes sense to bin + tweak those ddr4 chippies, slap on a pretty heatsink + obligatory leds, and make some more money. When DDR4 joins the ranks of tech where those doing the production run ask "Why are we still doing this? It's almost too easy.", well then you can get those DDR4@2400 modules given away for free with a packet of Wheaty Flakes. But not today.... Today you still have to deal with marketing bullshit and accept the fact that a given XMP profile may or may not work on that shiny new motherboard.

PS: Link to that similar ddr3 situation years ago: http://www.tomshardware.co.uk/forum/291857-30-tomshardware
« Last Edit: March 23, 2017, 11:35:04 am by mrflibble »
 

Offline TiN

  • Super Contributor
  • ***
  • Posts: 4543
  • Country: ua
    • xDevs.com
Re: DDR4 DIMM Lies
« Reply #26 on: March 23, 2017, 11:43:48 am »
:). Non-JEDEC DDR4 speed rated sticks are usually just prebinned, preselected modules, which tested to run at higher speeds in specific boards. XMP is one of the ways how to tell specific motherboard about capabilities of specific module, so BIOS can set timings, voltages and other related MC settings accordingly to make the module work. There can be a lot of things happening under the hood, outside of typical RAS-CAS-CMD timings by application of XMP profile.

That's why vendors of both memory sticks and mainboards publish so called AVL's, which tell you how well can DIMM A work in MB B, and what CPU was used, etc. This usually works quite well with Intel platforms. With Ryzen....well, you into some fun, I've heard. Personally I didn't try or touch anything AMD since very long time, but info above is based on many generations of Intel designs I was working on.  :)
YouTube | Metrology IRC Chat room | Let's share T&M documentation? Upload! No upload limits for firmwares, photos, files.
 

Offline mrflibble

  • Super Contributor
  • ***
  • Posts: 2051
  • Country: nl
Re: DDR4 DIMM Lies
« Reply #27 on: March 23, 2017, 11:49:02 am »
I don't know if ASrock has become better the last few years, but they've built cheap mainboards, cheap in all regards.
From what I hear they've kept up the fine tradition of "Cheap in all regards!". ;)

But then, every manufacturer produces a "suboptimal" product every now and then. So it's usually worth it to do your homework and go over the reviews.
 

Offline Mr. Scram

  • Super Contributor
  • ***
  • Posts: 9810
  • Country: 00
  • Display aficionado
Re: DDR4 DIMM Lies
« Reply #28 on: March 23, 2017, 01:31:52 pm »
Well of course they bin them - the point is that they're upselling a product they don't even manufacture. And they regularly do it by cranking the voltage past normal levels.
What's so bad about that? They test the modules, rate them according to their own findings and take responsibility for the consequences. That is not much different from the manufacturer doing the same thing. They put in effort and take a risk, which is what they charge you for. Video card manufacturers do the same thing with factory overclocked GPUs.

One might argue that they might have been clearer on this, but we have to realise that Ryzen was only released a week or so ago. The chips were apparently fully compatible with DDR4 systems on the market at that time, but Ryzen requires different qualities and therefore the testing process needs to be adjusted.

It is good to note that QVLs (Qualified Vendor List) are still a thing. Even if you have memory that complies to every standard right from the factory and use it in a random motherboard that does the same, operation is not guaranteed. It typically tend to work just fine, but some combinations throw a fit, even though the individual parts are adhering to standards and work fine in every other combination. The immense complexity of the hardware ecosystem has many benefits, but the one downside is that unexpected behaviour is an option.


I don't know if ASrock has become better the last few years, but they've built cheap mainboards, cheap in all regards.
I have seen a lot of systems being built with Asrock motherboards and there do not appear to be much problems. Performance seems to be decent to fairly good too.

The last few big motherboard issues came from Asus (finicky memory and heating issues), Gigabyte (cold boot issues) and, a while back, Intel itself (SATA chipset issues). Oh, and I guess we can count the Atom issues too, since those chips are soldered to the motherboard, but that might admittedly be a stretch.
« Last Edit: March 23, 2017, 01:38:55 pm by Mr. Scram »
 

Offline gnifTopic starter

  • Administrator
  • *****
  • Posts: 1672
  • Country: au
Re: DDR4 DIMM Lies
« Reply #29 on: March 23, 2017, 04:43:24 pm »
In the past I have gone with ASUS boards, but I have found that their support over time is terrible, getting BIOS updates with later microcode is a nightmare, it had left many of us to hack/patch our own BIOS images with updated microcode. So far ASRock I have noted have been releasing not just updated microcode, but new features, settings, etc... even for the older boards. They were one of the few that actually bothered to add the Phenom AM2+ support to their older boards.

As for the build quality of this board, it is pretty decent, I wouldn't say its awesome though, the heat sinks on the mosfets are pretty cheap and nasty, fixed with just those plastic spring things.

The issue here isn't that RAM is binned, etc... the issue is that they do not market the RAM at it's real clock, in many cases they don't even specify its 'default stable clock', ie, the real factory clock. So when you are shopping for RAM, unless you do what I was unaware of prior (buy from Crucial), you will be gambling on what you get.

One other thing to note, Micron do sell RAM rated at 1600MHz (DDR4-3200), according to digikey the ICs are around $13.70 USD each. 16 would be needed to make a 8GB DIMM, which would cost $219.20 just for the modules, making a 16GB pack cost $438.4. So unless you are paying that kind of cash for a DDR4-3200 set, you're likely getting overclocked junk.
 

Offline wraper

  • Supporter
  • ****
  • Posts: 16796
  • Country: lv
Re: DDR4 DIMM Lies
« Reply #30 on: March 23, 2017, 06:05:17 pm »
Easiest way to get RAM with true JEDEC speed, which don't need overclock/overvolt to work at advertised speed, is buying cheap ram without heatsinks  :palm:, like Kingston value RAM. The fun part is that cheap ram often have higher JEDEC speed than those more expensive "faster" types with fancy heatsinks.
 

Offline gnifTopic starter

  • Administrator
  • *****
  • Posts: 1672
  • Country: au
Re: DDR4 DIMM Lies
« Reply #31 on: March 23, 2017, 06:07:47 pm »
Easiest way to get RAM with true JEDEC speed, which don't need overclock/overvolt to work at advertised speed, is buying cheap ram without heatsinks  :palm:, like Kingston value RAM. The fun part is that cheap ram often have higher JEDEC speed than those more expensive "faster" types with fancy heatsinks.

This is what I did initially, I bought ram without any heatsinks, value/cheap ram, but it was programmed with overclocked SPD timings, no XMP timings of any sort.
 

Offline wraper

  • Supporter
  • ****
  • Posts: 16796
  • Country: lv
Re: DDR4 DIMM Lies
« Reply #32 on: March 23, 2017, 06:27:37 pm »
Easiest way to get RAM with true JEDEC speed, which don't need overclock/overvolt to work at advertised speed, is buying cheap ram without heatsinks  :palm:, like Kingston value RAM. The fun part is that cheap ram often have higher JEDEC speed than those more expensive "faster" types with fancy heatsinks.

This is what I did initially, I bought ram without any heatsinks, value/cheap ram, but it was programmed with overclocked SPD timings, no XMP timings of any sort.
Well, you wrote XMP  :-//:
Quote
Later I found out that these modules are JEDEC-2133 but XMP-2400 (Intel's Extreme Memory Profile), which is why they are selling them as DDR4-2400
Never heard about Avexir, also they are black an called platinum, so not really non fancy "value" stuff where buyer don't care how they look.
 

Offline Mr. Scram

  • Super Contributor
  • ***
  • Posts: 9810
  • Country: 00
  • Display aficionado
Re: DDR4 DIMM Lies
« Reply #33 on: March 23, 2017, 06:30:30 pm »
The issue here isn't that RAM is binned, etc... the issue is that they do not market the RAM at it's real clock, in many cases they don't even specify its 'default stable clock', ie, the real factory clock. So when you are shopping for RAM, unless you do what I was unaware of prior (buy from Crucial), you will be gambling on what you get.

That is the point: there is no such thing as a real clock. The chips get binned by their manufacturer, and then get binned again. Binning is just testing a chip against a certain speed and rating. There is no reason the second bin is less valid than the first. The first manufacturer does not guarantee anything beyond the lower memory clocks. The second manufacturer takes a longer look at the chips and guarantees a higher clock.

Let's put it this way: I sell you a hundred oranges that weigh at least 150 grams each. I put a sticker on them: at least 150 grams. You take these oranges and sort them according to weight, and put the 35 oranges that turn out to weigh more than 200 grams in some nice paper. You sell these fifth kilo oranges to people for a bit more money than the other oranges. What would you say to someone coming back with an orange, pointing at the sticker that these are actually 150 gram oranges?

Of course, due to compatibility issues, the situation with memory is a bit more complex than it is with oranges. That would amount to something like conventional scales that weigh the oranges just fine and a totally new model of scale that does not know how to deal with the oranges, but I digress. The binning part is pretty much the same. Adding Ryzen to the mix just means the binning requirements need to be a bit tighter than it was before.
 

Offline Monkeh

  • Super Contributor
  • ***
  • Posts: 7990
  • Country: gb
Re: DDR4 DIMM Lies
« Reply #34 on: March 23, 2017, 06:47:59 pm »
The issue here isn't that RAM is binned, etc... the issue is that they do not market the RAM at it's real clock, in many cases they don't even specify its 'default stable clock', ie, the real factory clock. So when you are shopping for RAM, unless you do what I was unaware of prior (buy from Crucial), you will be gambling on what you get.

That is the point: there is no such thing as a real clock. The chips get binned by their manufacturer, and then get binned again. Binning is just testing a chip against a certain speed and rating. There is no reason the second bin is less valid than the first. The first manufacturer does not guarantee anything beyond the lower memory clocks. The second manufacturer takes a longer look at the chips and guarantees a higher clock.

Guarantees a higher clock - so long as your board manages to hit weird timing specs, and cranks the voltage up outside of normal specs.

Use of XMP should be very, very clearly noted along with the actual, standard profile the RAM will operate at. Anything outside JEDEC profiles is 'best of luck to you'.
 

Offline Mr. Scram

  • Super Contributor
  • ***
  • Posts: 9810
  • Country: 00
  • Display aficionado
Re: DDR4 DIMM Lies
« Reply #35 on: March 23, 2017, 06:59:39 pm »
Guarantees a higher clock - so long as your board manages to hit weird timing specs, and cranks the voltage up outside of normal specs.

Use of XMP should be very, very clearly noted along with the actual, standard profile the RAM will operate at. Anything outside JEDEC profiles is 'best of luck to you'.
That is what the aforementioned QVL is for. Even memory that has been binned from the factory for a certain speed is not guaranteed to work at that speed with all motherboards. Only the memory on the QVL carries that guarantee of compatibility :)
 

Offline Monkeh

  • Super Contributor
  • ***
  • Posts: 7990
  • Country: gb
Re: DDR4 DIMM Lies
« Reply #36 on: March 23, 2017, 07:02:30 pm »
Guarantees a higher clock - so long as your board manages to hit weird timing specs, and cranks the voltage up outside of normal specs.

Use of XMP should be very, very clearly noted along with the actual, standard profile the RAM will operate at. Anything outside JEDEC profiles is 'best of luck to you'.
That is what the aforementioned QVL is for. Even memory that has been binned from the factory for a certain speed is not guaranteed to work at that speed with all motherboards. Only the memory on the QVL carries that guarantee of compatibility :)

The QVL is a crutch for poorly designed systems.

If you get a properly built module specified to and supplying JEDEC standard timings, it should function with any system which supports those speeds. If you do not support those speeds, then do not try and run them - pick the lower profile. If you 'support' those speeds and it doesn't work with a standard module: Your board or firmware is broken. Fix it.

There is no point in these standards if they're treated as optional.

Now, if there is some freak aberration which causes one or two very specific modules not to work - then you can write a 'sorry, this doesn't work' list, and I'll only think slightly less of you.
 

Offline Mr. Scram

  • Super Contributor
  • ***
  • Posts: 9810
  • Country: 00
  • Display aficionado
Re: DDR4 DIMM Lies
« Reply #37 on: March 23, 2017, 07:18:13 pm »
There is no point in these standards if they're treated as optional.

Now, if there is some freak aberration which causes one or two very specific modules not to work - then you can write a 'sorry, this doesn't work' list, and I'll only think slightly less of you.
I don't think they are treated as optional. It is a market with very complex products, technological capabilities that change every day and a nauseatingly fast development cycle. There are bound to be a number of hiccups. Just look at the errata of singular products that are much simpler and have a much slower development cycle. You quickly realise that it is a miracle that so many of these components just happen to be compatible at all.

The few issues that arise really do not warrant slowing the cycle way down and taking many years for what currently takes one.
 

Offline Someone

  • Super Contributor
  • ***
  • Posts: 4510
  • Country: au
    • send complaints here
Re: DDR4 DIMM Lies
« Reply #38 on: March 23, 2017, 10:36:48 pm »
There is no point in these standards if they're treated as optional.

Now, if there is some freak aberration which causes one or two very specific modules not to work - then you can write a 'sorry, this doesn't work' list, and I'll only think slightly less of you.
I don't think they are treated as optional. It is a market with very complex products, technological capabilities that change every day and a nauseatingly fast development cycle. There are bound to be a number of hiccups. Just look at the errata of singular products that are much simpler and have a much slower development cycle. You quickly realise that it is a miracle that so many of these components just happen to be compatible at all.

The few issues that arise really do not warrant slowing the cycle way down and taking many years for what currently takes one.
Its also a very competitive market where any small advantage can get huge profits, so there is a business case to squeeze out every last bit of timing. It falls down when that timing is against the dominant platform rather than the "guaranteed" jedec timings against all possible memory controllers. Just a bit of crappy marketing which seems to have been blown out of proportion rather than discussing the electronics side of it with timing/noise margins and the obvious differences between intel and AMD parts.
 

Offline Yansi

  • Super Contributor
  • ***
  • Posts: 3893
  • Country: 00
  • STM32, STM8, AVR, 8051
Re: DDR4 DIMM Lies
« Reply #39 on: March 23, 2017, 11:12:11 pm »
Easiest way to get RAM with true JEDEC speed, which don't need overclock/overvolt to work at advertised speed, is buying cheap ram without heatsinks  :palm:, like Kingston value RAM. The fun part is that cheap ram often have higher JEDEC speed than those more expensive "faster" types with fancy heatsinks.

I usually not poke my fingers inside PCs, as i really hate that PC stuff...  Anyway, what you said is interesting. As it was exactly what I have seen few months ago, when I wanted to upgrade RAM in my notebook. Just browsed local retailers eshops and quickly found out the cheapest RAMs did indeed had if not the best performace, regarding timing and such. I couldn't believe that too   :o
 


Share me

Digg  Facebook  SlashDot  Delicious  Technorati  Twitter  Google  Yahoo
Smf