Gentoo Forums
Gentoo Forums
Gentoo Forums
Quick Search: in
AMDFX-8350+ASUSM5A97R2 MemoryCompatibility(SolvedNotClosed)
View unanswered posts
View posts from last 24 hours

Goto page Previous  1, 2, 3  Next  
Reply to topic    Gentoo Forums Forum Index Kernel & Hardware
View previous topic :: View next topic  
Author Message
wcg
Guru
Guru


Joined: 06 Jan 2009
Posts: 588

PostPosted: Tue Apr 30, 2013 2:14 am    Post subject: Reply with quote

If you are using dual-channel memory, anything over 1333mhz is overclocked
for an AMD FX cpu: http://support.amd.com/us/kbarticles/Pages/ddr3memoryfrequencyguide.aspx
(Voids the warranty, if I read the various links from that page correctly.)

I bought an M5A97, Corsair Vengeance 1866, and a Phenom II 965.
There were various problems with the motherboard (PCIe slots and
a PCIe 1.0 x16 video card scrambled the BIOS setup display,
inability to clear CMOS, the sucky ethernet chip, etc.) I ended up
RMAing it within a week and replaced it with an ASRock 990FX Extreme4,
which was compatibile with the same cpu and ram. I kept the Phenom II cpu
and Corsair 1866 ram. The ram passes memtest86 with no problems
in the ASRock board:
http://www.asrock.com/mb/AMD/990FX%20Extreme4/?cat=Specifications
(If you look at the AMD cpu-memory speed chart at the URL above,
it says 1066 is spec for a Phenom II with dual-channel memory, so that
is what it is running at. No overclocking as long as the cpu is still
under warranty. Plus it is fast enough to not be annoying at stock speeds,
and I care more about stability than peak peformance.)

That Asrock board does have a few quirks: the BIOS lies about PCIe 1.0
incompatibility when I have a PCIe 1.0 video card in the first PCIe x16
slot (maybe any of them). If I install a PCIe 1.0 video card, it gives the
AMI BIOS beep code for no video (5 beeps), and there is no post display.
But if you enter BIOS setup, the BIOS screens display just fine, it displays
the grub bootloader menu at boot, and the kernel has no problems
with the PCIe 1.0 x16 video card, in either text mode or X.

If you replace the PCIe 1.0 x16 video card with a PCIe 2.0 x16 video card,
the beeps go away, you can see the post display, and everything else
is the same. (The PCIe 2.0 card is a little faster, of course, but none of
the software that I use was hampered by the performance of the PCIe
1.0 card.)

dmesg reports one quirk:
Code:

QUIRK: Enable AMD PLL fix


I do not know whether that is a cpu quirk or a 990FX chipset quirk.
(One day I will dig into the kernel code, find that message, and find out.)

So, based on my experience, your problems are more likely with
the M5A97 than with the Corsair 1866 ram.
_________________
TIA
Back to top
View user's profile Send private message
nlsa8z6zoz7lyih3ap
Guru
Guru


Joined: 25 Sep 2007
Posts: 388
Location: Canada

PostPosted: Fri May 03, 2013 2:30 am    Post subject: Reply with quote

I am starting to suspect that I don't even understand how to read the QVL memory list. For ex:

Quote:
CMZ32GX3M4X1866C10(Ver3.23)(XMP)
M5A97_R20-Memory-QVL says 4 DIMM not supportet only with 2


Does this mean that even 'tho it is listed in the QVL as 32 GB of memory, the QVL is saying that you can only plug 2 of the 4 chips into the mob and thus use only 16GB of the 32?
That never occured to me, but now it seems as if that is what it is saying, and that I erred in purchasing this memory in the first place. Many thanks for pointing this out to me as I missed it, but would you be kind enough to confirm that this is true?

Regarding:
Quote:
What was interesting is that for this category of motherboard, "max 32G", when I tried to get ECC DIMMs they would not recommend 32G, but only 24G. I'm presuming that from a sheer numeric point of view, 32G might be right at the top end in terms of bus loading, and when you start adding the extra chips for ECC you go over the limit


I need some educating here When you don't see an exact mob on the crucial memory selector web page, how does one decide what "class" an mob belongs to.
Beyond this, I am starting to get the impression that not all mob's with the same written specs will perform as desired. If stability and performance to spec are important, how does one make a good choice?

Also
Quote:
If you are using dual-channel memory, anything over 1333mhz is overclocked
for an AMD FX cpu: http://support.amd.com/us/kbarticles/Pages/ddr3memoryfrequencyguide.aspx
(Voids the warranty, if I read the various links from that page correctly.)


Thanks very much for this. Again I am beginning to realize how poorly I understood memory specs when I chose the memory for my new PC. Your reference and comments, and those made by others will be the basis of my attempt to read and fill in my deficiencies in this area.



Regarding my question:
Quote:
Question: What is the correct kernel setting for cpu type. ? Opteron/Athlon64/Hammer/K8 comes with the advice Select this for an AMD Opteron or Athlon64 Hammer-family processor. │
│ Enables use of some extended instructions, and passes appropriate │


I have now compiled my world with gcc-4.7.2 with -march=native . After reading https://bbs.archlinux.org/viewtopic.php?id=154333, I patched my kernel to natively support this cpu type.
Everything works stably and just fine.
Back to top
View user's profile Send private message
depontius
Advocate
Advocate


Joined: 05 May 2004
Posts: 3509

PostPosted: Fri May 03, 2013 10:52 am    Post subject: Reply with quote

Now to blur matters even more, toss more information into the confusion ring, etc...

I suspect the real issue here is chip density. The drivers on the Vishera can drive so much capacitance. Some of that capacitance is in the sockets, some is in the board traces, but what we're concerned about is the memory chip inputs. Essentially the CPU can drive only so many memory chips.

I was not surprised to see what I did about the ECC, because ECC uses 1/9 more memory chips to store the syndrome bits.

The real question becomes, does the 8G DIMM use more chips, or denser chips than the 4G DIMM. If it uses more chips, then the capacitive load is building up, even though it's only 2 DIMMs.

For my part, I need to email Crucial, or someone. Because 24G can be done with 3 8G DIMMs, or with 2 8G and 2 4G. Because pairs perform better, I would expect the latter to perform better, unless there's some sort of secret sauce in there for triplets.

This may also be a moving target, because next year there may be denser memory chips, meaning fewer needed on an 8G DIMM, and then more could be used. Typically the input capacitance is constant, regardless of density.

(As a matter of fact, I AM a chip designer.)
_________________
.sigs waste space and bandwidth
Back to top
View user's profile Send private message
_______0
Guru
Guru


Joined: 15 Oct 2012
Posts: 521

PostPosted: Tue May 07, 2013 2:22 pm    Post subject: Reply with quote

depontius wrote:
Now to blur matters even more, toss more information into the confusion ring, etc...

I suspect the real issue here is chip density. The drivers on the Vishera can drive so much capacitance. Some of that capacitance is in the sockets, some is in the board traces, but what we're concerned about is the memory chip inputs. Essentially the CPU can drive only so many memory chips.

I was not surprised to see what I did about the ECC, because ECC uses 1/9 more memory chips to store the syndrome bits.

The real question becomes, does the 8G DIMM use more chips, or denser chips than the 4G DIMM. If it uses more chips, then the capacitive load is building up, even though it's only 2 DIMMs.

For my part, I need to email Crucial, or someone. Because 24G can be done with 3 8G DIMMs, or with 2 8G and 2 4G. Because pairs perform better, I would expect the latter to perform better, unless there's some sort of secret sauce in there for triplets.

This may also be a moving target, because next year there may be denser memory chips, meaning fewer needed on an 8G DIMM, and then more could be used. Typically the input capacitance is constant, regardless of density.

(As a matter of fact, I AM a chip designer.)
Back to top
View user's profile Send private message
_______0
Guru
Guru


Joined: 15 Oct 2012
Posts: 521

PostPosted: Tue May 07, 2013 2:25 pm    Post subject: Reply with quote

depontius wrote:
Now to blur matters even more, toss more information into the confusion ring, etc...

I suspect the real issue here is chip density. The drivers on the Vishera can drive so much capacitance. Some of that capacitance is in the sockets, some is in the board traces, but what we're concerned about is the memory chip inputs. Essentially the CPU can drive only so many memory chips.

I was not surprised to see what I did about the ECC, because ECC uses 1/9 more memory chips to store the syndrome bits.

The real question becomes, does the 8G DIMM use more chips, or denser chips than the 4G DIMM. If it uses more chips, then the capacitive load is building up, even though it's only 2 DIMMs.

For my part, I need to email Crucial, or someone. Because 24G can be done with 3 8G DIMMs, or with 2 8G and 2 4G. Because pairs perform better, I would expect the latter to perform better, unless there's some sort of secret sauce in there for triplets.

This may also be a moving target, because next year there may be denser memory chips, meaning fewer needed on an 8G DIMM, and then more could be used. Typically the input capacitance is constant, regardless of density.

(As a matter of fact, I AM a chip designer.)


non-sense, industry made sure all pieces work together and mobos advertising 32GB are supposed to work with 32GB ram. Bios stock settings should make things work. If the bios has some crazy OC section with custom values disable all or set everything to AUTO

Another trick, try to change the mem modules around.
Back to top
View user's profile Send private message
nlsa8z6zoz7lyih3ap
Guru
Guru


Joined: 25 Sep 2007
Posts: 388
Location: Canada

PostPosted: Tue May 07, 2013 4:50 pm    Post subject: Reply with quote

Quote:
I suspect the real issue here is chip density. The drivers on the Vishera can drive so much capacitance. Some of that capacitance is in the sockets, some is in the board traces, but what we're concerned about is the memory chip inputs. Essentially the CPU can drive only so many memory chips.


Thanks so much for this. Now I feel that I understand what is going on. I used to make Heathkit electronics in the 1950's. In those analog days concepts such as capacitance and inductance
played the key roles; moreover every component in (say) a shortwave radio interacted with everything else in the kit;--the internals of one component were not entirely isolated from the others.
When I started choosing parts and assembling PC's about 8 years ago, I did realize that "the real world of physics" occurred inside each component , but believed (as _______0 seems to in the post above) that each component in a computer interacts with the others in a purely perfect and exact digital fashion, that the actual physics inside one component is totally hidden from all of the others. I now no longer believe that this is true.



wcg gave us a reference from AMD which is well worth reading carefully and in its entirety. (http://support.amd.com/us/kbarticles/Pages/ddr3memoryfrequencyguide.aspx).

Now all of this does make the personal selection of components for a home built PC much harder than I had realized before building this one, but perhaps not impossible. Armed with depontius's insights, I would approach the part selection quite differently than I did before, specifically checking with the cpu manufacturer's info as well as the mob's, before choosing memory.

Regarding:

Quote:
Bios stock settings should make things work. If the bios has some crazy OC section with custom values disable all or set everything to AUTO


Very good advice. However all of my problems occurred using non-overclocked stock settings.

Quote:
Another trick, try to change the mem modules around


This is also very good advice. However the 2nd set of modules was tried individually (one at a time) and to-gether as sets of 3 and 4 in several different orders while doing my testing.

Regarding:

Quote:
mobos advertising 32GB are supposed to work with 32GB ram
This may well be true. However as trismo kindly pointed out:

Quote:
M5A97_R20-Memory-QVL says 4 DIMM not supportet only with 2


So I really goofed in my memory choice for this PC. Although this memory comes with 4 DIMM, the mob QVL only claims that it is supported when 2 of them are used.
Curious that I could have spent so much time selecting parts and still missed something so important just because I was not looking for it.

Thanks again to everyone for their postings. I am really enjoying and benefiting from this forum topic, and encourage people to continue replying with their thoughts and knowledge regarding it.
Back to top
View user's profile Send private message
depontius
Advocate
Advocate


Joined: 05 May 2004
Posts: 3509

PostPosted: Tue May 07, 2013 5:07 pm    Post subject: Reply with quote

_______0 wrote:

non-sense, industry made sure all pieces work together


Bought any bridges lately? I agree that the pieces all should work together, and stock BIOS settings would be the most reliable way to go. But remember that most of these motherboards aren't engineered, they're thrown together based on the reference design(s) from the chip supplier, with some level of testing. The amount of engineering and testing will vary widely. The profit in the PC industry is nearly all focused on Microsoft software and Intel processors. There are probably some other reasonably profitable components, as well. The rest pretty much runs on a shoestring. If you want an engineered motherboard, get a server motherboard. Otherwise, as I said, you're getting tweaks of reference design(s).
_________________
.sigs waste space and bandwidth


Last edited by depontius on Tue May 07, 2013 9:49 pm; edited 1 time in total
Back to top
View user's profile Send private message
_______0
Guru
Guru


Joined: 15 Oct 2012
Posts: 521

PostPosted: Tue May 07, 2013 9:14 pm    Post subject: Reply with quote

http://support.amd.com/us/kbarticles/Pages/ddr3memoryfrequencyguide.aspx

^^^ interesting link. Good in order not to, blindly, get random stuff.

What about overclocking the RAM? I've read that's a trivial OC to easily get higher mem clock. Though I still have to try it.
Back to top
View user's profile Send private message
nlsa8z6zoz7lyih3ap
Guru
Guru


Joined: 25 Sep 2007
Posts: 388
Location: Canada

PostPosted: Tue May 07, 2013 10:31 pm    Post subject: Reply with quote

Quote:
What about overclocking the RAM? I've read that's a trivial OC to easily get higher mem clock. Though I still have to try it.

The memory I am used is rated by Corsaire as 1866, but according to the amd link should run at 1333 with the FX-8350 cpu. (And of course I am using 3 of the 4 dimms whereas the QVL for the mob only indicates that 2 of the 4 can be used. So in that sense anything over 1333 is overclocked.

While I am not an overclocker, I couldn't help playing with the bios settings. My partiial results are that the following run stably.

(1) Default CPU settings with 1600 for RAM with 1.35volts
(2) Using the mob overclock tuner to set the cpu to 4320 and memory to 1439 with 1.35 volts.

As my primary interest is in a fast multiprocess stable PC, I haven't tried much else.

Maybe as an overclocker you would know the answer to this question: Does over-clocking the memory actually increase performance?
Back to top
View user's profile Send private message
_______0
Guru
Guru


Joined: 15 Oct 2012
Posts: 521

PostPosted: Tue May 07, 2013 10:52 pm    Post subject: Reply with quote

it does, and orders of magnitude.

check one of my posts back where I am testing my ram speed and somebody else does the same test but double the transfer rate. I think my post was about slow rsyncing within ram.
Back to top
View user's profile Send private message
depontius
Advocate
Advocate


Joined: 05 May 2004
Posts: 3509

PostPosted: Wed May 08, 2013 1:14 am    Post subject: Reply with quote

Going back to my post I mentioned "denser chips or more chips." Looking at the AMD page cited, look in colums 4, 5, and 6 and you'll see some of what I was talking about, but wasn't using the right language. "Single rank memory" means essentially that there are just enough chips to fill the 64-bit data bus. To be honest, I don't know what configuration "standard density" is these days, so I can't tell you how many chips that comes to. But "double rank memory" means that there are twice that many chips, and you select one rank or the other. Double rank also means roughly twice the input capacitance on almost all signal pins.

ECC DIMMs have a 72-bit data bus instead of 64-bits. That means 11% more chips on the DIMM. The extra capacitance (specifically extra capacitance on the address pins) of the extra chips is why I supposed Crucial only recommended 24G on a "32G" system. The part where I talked of using 2 8G and 2 4G DIMMs to perhaps get better performance in a 24G system would be their column 6.

Look a little further down in the "Crosshair V Formula" table and you'll see the SS/DS field. I'm guessing that that means "Single-Sided" and "Double-Sided", meaning chips on one side of the DIMM or both sides. At least in the past this typically corresponded to single rank and double rank, and probably does still.

The last standalone DRAM I worked on was a 16Mb SDRAM. Since then I've been working in embedded DRAM.
_________________
.sigs waste space and bandwidth
Back to top
View user's profile Send private message
Anon-E-moose
Watchman
Watchman


Joined: 23 May 2008
Posts: 6095
Location: Dallas area

PostPosted: Wed May 08, 2013 9:14 am    Post subject: Reply with quote

nlsa8z6zoz7lyih3ap wrote:
(1) Default CPU settings with 1600 for RAM with 1.35volts


Why 1.35 volts, standard voltage is 1.5, and perhaps up to 1.55 to 1.6 for overclocking.

With that much undervolting, I would think even running at 1333 would be problematic.

You could try manually setting the memory timings to higher than spec and see if it becomes stable also.

at 1333 speed the timing should be 9-9-9-24
at 1866 it increases to 10-11-10-30

You might manually set them to 10-11-10-30 and see what they do at 1600
and it could be that the memory controller just won't handle 1600 and up speed.
_________________
PRIME x570-pro, 3700x, 6.1 zen kernel
gcc 13, profile 17.0 (custom bare multilib), openrc, wayland
Back to top
View user's profile Send private message
wcg
Guru
Guru


Joined: 06 Jan 2009
Posts: 588

PostPosted: Wed May 08, 2013 12:42 pm    Post subject: Reply with quote

So "dual rank" and "dual channel" mean different things? Interesting.

I will note that both the RMAed M5A97 and the ASRock 990FX Extreme4
autotuned the memory to 1333mhz in dual channel mode, and that
was the speed it was set at when testing with memtest86. (I downclocked
it to 1066 manually.) The Corsair 1866 was not on the QVL for the ASRock
board, but I suspected that it would work, since other speeds of Corsair
memory were there, and in fact it does work fine on that mb.
_________________
TIA
Back to top
View user's profile Send private message
depontius
Advocate
Advocate


Joined: 05 May 2004
Posts: 3509

PostPosted: Wed May 08, 2013 1:21 pm    Post subject: Reply with quote

wcg wrote:
So "dual rank" and "dual channel" mean different things? Interesting.


The data path width for a single DIMM is 64 bits - 72 bits (I think, it might be more.) for an ECC DIMM. For "dual channel" the data path width is double that - 128 bits for the non-ECC DIMMs. That doubles the memory bandwidth, and that is why "dual channel" is advertised as the higher-performance method. As previously mentioned, "dual rank" essentially means two sets of chips on a DIMM, so you can get data out of one set or the other. It doubles your memory density, but at the same performance.
_________________
.sigs waste space and bandwidth
Back to top
View user's profile Send private message
wcg
Guru
Guru


Joined: 06 Jan 2009
Posts: 588

PostPosted: Wed May 08, 2013 10:57 pm    Post subject: Reply with quote

Quote:

Why 1.35 volts, standard voltage is 1.5, and perhaps up to 1.55 to 1.6 for overclocking.


Some dimms are rated for 1.35 volts. Examples:

Desktop:
http://www.newegg.com/Product/Product.aspx?Item=N82E16820148660
http://www.newegg.com/Product/Product.aspx?Item=N82E16820233244

Laptop:
http://www.newegg.com/Product/Product.aspx?Item=N82E16820148614
http://www.newegg.com/Product/Product.aspx?Item=N82E16820226344

Server:
http://www.newegg.com/Product/Product.aspx?Item=N82E16820239478
http://www.newegg.com/Product/Product.aspx?Item=N82E16820148628
_________________
TIA
Back to top
View user's profile Send private message
Anon-E-moose
Watchman
Watchman


Joined: 23 May 2008
Posts: 6095
Location: Dallas area

PostPosted: Thu May 09, 2013 12:22 am    Post subject: Reply with quote

wcg wrote:
Quote:

Why 1.35 volts, standard voltage is 1.5, and perhaps up to 1.55 to 1.6 for overclocking.


Some dimms are rated for 1.35 volts. Examples:


The one he mentioned isn't.

Code:
    DDR3 1866
    Timing 10-11-10-30
    Cas Latency 10
    Voltage 1.5V


_________________
PRIME x570-pro, 3700x, 6.1 zen kernel
gcc 13, profile 17.0 (custom bare multilib), openrc, wayland
Back to top
View user's profile Send private message
wcg
Guru
Guru


Joined: 06 Jan 2009
Posts: 588

PostPosted: Thu May 09, 2013 6:36 am    Post subject: Reply with quote

Quote:
The one he mentioned isn't.


Yes, that could cause some issues, running 1.5volt memory at 1.35volt.
_________________
TIA
Back to top
View user's profile Send private message
wcg
Guru
Guru


Joined: 06 Jan 2009
Posts: 588

PostPosted: Thu May 09, 2013 11:27 am    Post subject: Reply with quote

I found this brief description of a "memory rank" on Wikipedia:
http://en.wikipedia.org/wiki/Memory_rank

(Attempt to see exactly what the AMD cpu table "dual rank" notation
was referring to. I can't say the picture is much clearer after reading
that and a bunch of related links there. I get the impression that
"everything is dual, quad, octal, etc rank these days", which would
make the AMD table nonsensical, as single rank dimms would be pretty
hard to find, and it does not even mention quad-rank, etc. But that is
only an impression from the way those documents are written, not
necessarily reflecting engineering reality.)

At least memtest86 tells you how wide a data path to ram your mb
is using, so you will know if you are really running in "dual channel"
mode, regardless of how the dimms are organized.

edit:
One poster on another page complained that the Wikipedia
"memory rank" page has more information on "how" than
"why" (dual ranks on the same dimm cannot be accessed
simultaneously, as they share the same data path; it is not
equivalent to "dual channel", which aggregates two N-bit
data paths to make a 2N-bit data path).
_________________
TIA
Back to top
View user's profile Send private message
nlsa8z6zoz7lyih3ap
Guru
Guru


Joined: 25 Sep 2007
Posts: 388
Location: Canada

PostPosted: Sun May 12, 2013 5:48 pm    Post subject: Reply with quote

Performance with Different memory speed configurations

I ran some tests to see how memory speed affects the sort of applications that I actually run on my PC.
Tests were run with the mob overclock tuner and the choices of memory speed that it allows.

Test 1 x264 encoding
(i) 1728mghz memory
encoded 182304 frames, 20.51 fps, 22095.24 kb/s

real 148m9.750s
user 1157m45.780s
sys 1m45.266s

(ii) 864mghz memory
encoded 182304 frames, 20.04 fps, 22095.24 kb/s

real 151m38.255s
user 1182m20.954s
sys 2m7.435s

Comment: In my scheme of things differences of a few % are insignificant. Thus fast memory does not appear to be important for this task.

Test 2 My own programs concerning certain finite algebraic structures. Only type integer and operations mod 2 are used here. Structures are created in stages and stored in memory.
When new ones are created, it is necessary to retrieve the previously created ones from memory and compare them to the one currently being created.


(i) 1728 mghz memory and 8 instances of the program run simultaneously and independently with the same input
real 70m30.381s
user 69m44.433s
sys 0m1.519s

real 70m32.613s
user 69m46.014s
sys 0m1.544s

real 70m33.274s
user 69m45.189s
sys 0m1.456s

real 70m33.657s
user 69m44.019s
sys 0m1.484s

real 70m33.709s
user 69m44.682s
sys 0m1.478s

real 70m34.992s
user 69m43.341s
sys 0m1.494s

real 70m35.011s
user 69m43.182s
sys 0m1.471s

real 70m35.293s
user 69m44.040s
sys 0m1.447s

(ii) The same as the above but with only on instance run
real 53m18.922s
user 53m13.066s
sys 0m2.064s

Comment: We see quite strikingly that the 8 fx-8350 modules do not scale linearly whereas in previous 3GHZ 4 core phenom II they did. However even when 8 instances are run, they compare favourably with the phenom II 4 core which gave this result with 4 instances:
user 112m55.530s
sys 0m0.355s

real 114m8.204s
user 113m36.745s
sys 0m0.097s

real 114m36.089s
user 113m35.713s
sys 0m1.559s

real 114m54.924s
user 114m13.386s
sys 0m7.733s

Now back to the fx-8350

864 mghz memory and 8 instances of the program run simultaneously and independently with the same input

real 70m13.711s
user 69m46.721s
sys 0m2.614s

real 70m16.342s
user 69m47.318s
sys 0m2.674s

real 70m17.083s
user 69m44.463s
sys 0m2.638s

real 70m18.037s
user 69m46.672s
sys 0m2.672s

real 70m18.180s
user 69m45.355s
sys 0m2.556s

real 70m18.520s
user 69m45.227s
sys 0m2.475s

real 70m20.164s
user 69m46.312s
sys 0m2.528s

real 70m25.131s
user 69m44.602s
sys 0m2.549s

COMMENT: With this application, it is also true that memory speed has little impact on performance.

I did numerous other tests, including ones that varied the cpu speed, and concluded that for my important applications the bottleneck is cpu speed. I am not including the actual numbers from these additional tests here.

Of course, as has already been pointed out in this thread, there are other applications for which memory speed make a big difference.
Back to top
View user's profile Send private message
nlsa8z6zoz7lyih3ap
Guru
Guru


Joined: 25 Sep 2007
Posts: 388
Location: Canada

PostPosted: Sun May 12, 2013 6:15 pm    Post subject: Reply with quote

Quote:
Why 1.35 volts, standard voltage is 1.5, and perhaps up to 1.55 to 1.6 for overclocking.


This was because of matty's post earlier in this thread
Quote:
Setting DRAM voltage to the 1.35V(instead the 1.5V default) removed the memory failures with memtest and made the system stable. The downside of this setting is that it won't let you set RAM to more that 1600Mhz.


I can't give the reason why, but it works for me too. Might this have anything to do with capacitance and the fx-8350 having a "weak" memory controller?
Back to top
View user's profile Send private message
wcg
Guru
Guru


Joined: 06 Jan 2009
Posts: 588

PostPosted: Tue May 14, 2013 2:20 am    Post subject: Reply with quote

Quote:
can't give the reason why, but it works for me too. Might this have anything
to do with capacitance and the fx-8350 having a "weak" memory controller?


That is an interesting question. I have been using the Corsair 1866 right along
at 1.5v with the Phenom II on a 990FX chipset board, no problems. I would suspect
something with the 970 chipset mb or the BIOS for it with that new cpu rather than
with the cpu itself. (CPUs do have erratas, just not on the same scale that BIOS
versions, mb revisions, and chipsets have erratas. Marketing phrases like
"new soon-to-be-released cpu ready" always seem like rank speculation
rather than fact to me. We will not really know that it was "ready" until after
a few million users have tried it with that faster cpu. If it works, all good; if not,
hope that a BIOS upgrade can fix it.)
_________________
TIA
Back to top
View user's profile Send private message
Anon-E-moose
Watchman
Watchman


Joined: 23 May 2008
Posts: 6095
Location: Dallas area

PostPosted: Tue May 14, 2013 11:47 am    Post subject: Reply with quote

nlsa8z6zoz7lyih3ap wrote:
I can't give the reason why, but it works for me too. Might this have anything to do with capacitance and the fx-8350 having a "weak" memory controller?


As far as the cpu memory controller and the vishera cpus in general read http://www.tomshardware.com/forum/323498-30-need-8350

My personal opinion, speed differences between 1333, 1600 and higher don't matter much,
unless you're looking to either have spectacular benchmarks or uber-gaming performance
neither of which apply to most people with most computing needs.


As far as 1.35 vs 1.5 voltage, I would imagine that with 1.35 the timings become much looser (gives stability)
but the same could be done at 1.5 and not using auto memory timings.

I overclock my FSB which gives me greater cpu speed, but the memory drops back to 1066 speed
which is acceptable, to me, as the greater cpu speed makes up for the slightly slower memory access.

Bottom line do whatever works for you.
_________________
PRIME x570-pro, 3700x, 6.1 zen kernel
gcc 13, profile 17.0 (custom bare multilib), openrc, wayland
Back to top
View user's profile Send private message
depontius
Advocate
Advocate


Joined: 05 May 2004
Posts: 3509

PostPosted: Tue May 14, 2013 12:12 pm    Post subject: Reply with quote

Anon-E-moose wrote:

My personal opinion, speed differences between 1333, 1600 and higher don't matter much,
unless you're looking to either have spectacular benchmarks or uber-gaming performance
neither of which apply to most people with most computing needs.


Years back when I was building a system for my daughter's high-school graduation present, the prevailing wisdom seemed to be that you were best off running your memory at the FSB speed. You could run it faster, but then you'd have to "slip-sync" all of the signals moving them between the two clock domains. You would almost always lose some fraction of the higher-frequency clock interval, either grabbing controls/addresses/data from the FSB to pass to the memory, or lose some fraction of the higher-frequency clock passing data back to the FSB. Your savings might come because the higher-frequency memory clock could count the RAS and CAS latencies with finer granularity. (Hint - the underlying mechanisms of RAS and CAS latency are measured in nanoseconds, clock cycles are just what's available to measure that.) The chipset FSB for her mobo was 800MHz, and that's the frequency rating of DDR2 I got her.

By the way, as for that link, I don't believe there is such a thing as "quad channel", at least not on the Vishera. There is dual-channel, with 2 DIMMs per channel. If it were truly quad channel, I suspect that the speed rating wouldn't drop when going from dual to quad, because all of the signals would be separate, and the only common limitation would be the power distribution on the chip. (Though TDB-limited performance is becoming significant, these days.)
_________________
.sigs waste space and bandwidth
Back to top
View user's profile Send private message
Anon-E-moose
Watchman
Watchman


Joined: 23 May 2008
Posts: 6095
Location: Dallas area

PostPosted: Tue May 14, 2013 1:31 pm    Post subject: Reply with quote

depontius wrote:
By the way, as for that link, I don't believe there is such a thing as "quad channel", at least not on the Vishera. There is dual-channel, with 2 DIMMs per channel. If it were truly quad channel, I suspect that the speed rating wouldn't drop when going from dual to quad, because all of the signals would be separate, and the only common limitation would be the power distribution on the chip. (Though TDB-limited performance is becoming significant, these days.)


The vishera line was supposed to incorporate quad channel memory by adding a new memory controller.

As far as the speed drop, it would make a difference if quad channel was using a different memory controller than dual channel
and I don't know the particulars of it as I don't work at AMD, just stating an opinion.

Anyway, as I told him previously he should do what works for him, on his hardware. That's all anyone can do.


Edit to add: the memory he was using (referenced earlier in thread) is quad channel that can run as dual
but I don't know if there is a setting on the MB for one vs the other.
_________________
PRIME x570-pro, 3700x, 6.1 zen kernel
gcc 13, profile 17.0 (custom bare multilib), openrc, wayland
Back to top
View user's profile Send private message
wcg
Guru
Guru


Joined: 06 Jan 2009
Posts: 588

PostPosted: Tue May 14, 2013 1:46 pm    Post subject: Reply with quote

[edit:] (Different issues: quad rank is a potential internal dimm organization,
and I referenced it when I questioned it's omission in the AMD cpu
to memory rank/clock table; quad channel is a potential mb
dimm slot organization.) [/edit]

("quad rank", not "quad channel")

Dram power does matter, and that probably answers any questions about
whether the dimms on the mb QVLs for inexpensive desktop mbs are likely
to be single-rank or multi-rank dimms. I stumbled on this obliquely while
browsing for any recent comments on Vishera in comp.arch:

Quote:

With select DDR3 DRAM parts operating in a single-rank configuration
(point to point with no active write termination) it is possible to
get DRAM power efficiency for open page accesses in the range of
20 mW/Gbs. Multi-rank systems (requiring active read and write
termination) and closed-page access can increase this to over
100 mW/Gbs.

Taking this as a reasonable range of DRAM power efficiencies, your
0.5 TB/s would require a minimum of 80W of DRAM power and a maximum
of over 400W of DRAM power.


( http://newsgroups.derkeiler.com/Archive/Comp/comp.arch/2013-05/msg00052.html )

In other words, using multi-rank dimms may fit more memory on a dimm,
but it can require 5 times as much dram power as a single-rank dimm in
operation.

(It is not only a question of "is your power supply up to it". Is the power plane
on the mb up to it, and what does it take to cool it?)
_________________
TIA
Back to top
View user's profile Send private message
Display posts from previous:   
Reply to topic    Gentoo Forums Forum Index Kernel & Hardware All times are GMT
Goto page Previous  1, 2, 3  Next
Page 2 of 3

 
Jump to:  
You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot vote in polls in this forum