AMD CPU speculation... and expert conjecture

Page 460 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Toggle sidebar Toggle sidebar
Status
Not open for further replies.
It is weird then, that in practice it works like a charm. And also, why does Windows allows you to use 0 bytes of page file and even re-size it (on any drive you want).

How that works then? As long as I don't have a program that uses more than the RAM amount of page file memory it will be fine?

Are you sure there has been no change in how NT 6.0+ manages those malloc requests?

Cheers!
 


The page file is only actually used when your physical memory is full or there is more memalloc's made then physical memory allows for (NT will preemptively try to swap things to disk before a program writes to nearly full memory). Setting it do 0 size is the same as disabling it, which is not recommended. Having no pagefile means whenever a program does a memalloc there is nothing to reserve, NT will still work but there's no guarantee that memory will be always be available. Of course if you have significantly more physical memory then you'll ever use then there shouldn't ever be a time when the pagefile needs written to and you can safely disable it.

As for why the file grows and shrinks, that's because NT is resizing it to ensure there is enough space on the disk to guarantee those reservations. If your pagefile is 4GB, then grows to 8GB, that doesn't mean NT just wrote 4GB of data to your disk. All it did was allocate more empty NTFS blocks to the pagefile, essentially' it's reserving disk space for the file that reserves memory space. This was done because people would complain that the swapfile took up too much space on the HDD and it wasn't being used anyway. Notice there is a minimum and maximum setting, you can set 2GB minimum (required for kernel crash dumps) and 32GB maximum. What this does is when there is no memallocs it will be 2GB, which is large enough to fit the legacy core kernel memory address space, when program start requesting memory it will expand to ensure those reservations can be met. Since the kernel itself rarely exceeds ~400MB it won't expand until you've had enough memallocs to exceed the minimum size.

The key takeaway is that putting a pagefile on a ramdisk is very very dumb. Set it's minimum size to 2GB on the C: drive (for kernel crash dumps) and leave it alone. If your using a SSD and want to use a ramdisk for temp folders then put it on the HDD if you have one cause it'll want to balloon out unless you set it's maximum to 2GB, and then it'll complain about your system being almost out of memory. NT see's the pagefile as a guarantee, a backup, a failsafe, just in case programs attempt to utilize more memory then you have on the system. It's more concerned about "potentially" not about actual utilization.

I have 16GB of memory on my main system and once tried to make a 8GB ramdisk. I had my pagefile set to a fixed 2GB of size. Windows 7 x64 Ultimate complained constantly that I was running out of memory space even though I had 4GB of the remaining 8GB unallocated. It was panicking because it couldn't guarantee that the system would remain stable should another memory hungry application start up (ramdisk's are treated as a huge memory hungry program). This wouldn't be such an issue if MS built ramdisk support into NT's memory model like Unix/Linux has.
 



I wonder did techreport use yesterdays update for bf4, most are reporting the 'skip' to have disappeared, it was the game not mantle.

 


The relevant part of Rajeeb Hazra quote is when he says "except that". As I said before to you an homogeneous system uses LCUs and those don't run "highly parallel applications" in an optimal way. However the KL runs "highly parallel applications" as a TCU. This is why he adds "except that". The KL is not a LCU as I have been saying you for days.

I agree on that his quote is a bit confusing.* If you pay attention to the rest of the interview he also says that KL:

does something quite remarkable, and can be seen as an inflection point in the evolution of heterogeneous computing.
Click to expand...

As I said before, how Intel achieves heterogeneity (LCU+TCU) with the new KL is a mystery, because there is little data and some slides are contradictory. With data available today my hypothesis is that KL introduces something similar to BIG.little. The KL core has a cluster with OoO execution and a separate cluster with in-order execution.

I also mentioned you before that IBM is abandoning the homogeneous design in the IBlueGene/Q by a new heterogeneous architecture that is developing in collaboration with Nvidia. IBM will make the LCUs and Nvidia the TCUs


* The structure of his quote is similar to "I am always right, except when I make mistakes." You cannot take the part before "except" and pretend that the author is always right.



I have given you a link from February 2014 with the plans of all the foundries for 14nm. Only STMicro chose FDSOI, all the other foundries (including IBM) chose FinFETs.

The roadmap discussed in the SA link that I also gave you is the last roadmap from GF. There is no new roadmap that replaces that. You can see by yourself GF plans on their site. This is the link to their FINFET on bulk 14nm process

http://www.globalfoundries.com/technology/14XM.aspx

AMD is tapping out 20nm bulk (likely Carrizo) and then will migrate to FINFETs.
 
In regards to the Pagefile and the like, NT ALWAYS allocates RAM in terms of Virtual Memory anyway; all disabling the Pagefile does for you is forcing a 1:1 correlation between Virtual and Real addresses. You guarantee no paging occurs whatsoever, but things WILL break if you run out of addresses, since you can't page out address blocks that aren't in use.

In this day and age, there's no reason to mess with the pagefile at all.
 


We can see clearly Intel advantage on the CPU side and AMD advantage on the GPU side. We can also see memory bottleneck on AMD side.

I believe AMD released MANTLE to compete against Broadwell. Broadwell promises a big update on the GPU side (some reports mention up to 80% gains over Haswell GPU, other mention up to 40%) and even desktop K series will include Iris Pro level of integrated graphics.

If current low-power i5 R-Haswell is rather close to top Kaveri A10 on graphics, I believe that future full-desktop 'i5-5670k' could be faster than Kaveri and AMD would need to add MANTLE to the equation. I don't expect big gains for Carrizo GPU due to memory bottlenecks.

The good news is that Broadwell has been delayed again

http://www.electronicsweekly.com/news/general/intels-14nm-mobile-delayed-till-2015-2014-02/

and that Intel is losing its traditional process advantage over competitors. As mentioned above TMSC is accelerating its roadmaps to release 16nm FinFET this year.
 
Dual graphics review is up:

http://www.techspot.com/review/781-amd-a10-7850k-graphics-performance/

The final conclusion?

If possible, we recommend spending a little extra money on a competent mid-range graphics card for guaranteed performance in all of your favorite games. While AMD's latest APUs make an interesting value proposition on their own, it seems like there are few situations where it would make sense to buy the company's FM2+ platform for Dual Graphics between an APU and a low-end discrete GPU.
Click to expand...
 
Thanks for the links, palladin. Then I won't be using a RAMDisk alternative in Windows any time soon, haha. I have the Page File set to 2GB IIRC, but I have to check. So far, no program has crashed, but I've seen "low memory" warnings. I guess I'l put 4GB now, haha.

And in regards to dual graphics with APUs. It's always been a bad idea, but there's one advantage to it. If you really really need more GPU power, you still have the option to do it without making an expensive purchase. There's a zone where you'll need to change the whole thing, but it gives you a little more stretch to play with before reaching that point.

The only safe bet to HTPC upgrade-ability is to, for example, use a K series from Intel (low heat/power are king in SFF) and then swap the GPU accordingly. Problem is, past the low end spectrum, there's no SFF models for mid-range and up in GPUs. I had a hard time finding a 6670 GDDR5 in low profile (here in Chile, that is) and now I'm having a hard time finding a low profile 250X/260 GDDR5. If you need more GPU power with an Intel GPU, you're busted in low profiles. And to be honest, the new trend of bulky "SFF" is pretty crappy. The closest thing is the CM 361, which I'm debating with myself to buy or not, haha. But that's as good as it gets for "low profile" desktops. And seriously folks, take a look at that case if you're planning on any SFF buid.

Anyway, like in every build, the thing is you have to take into account what your long term plan is. For a low profile desktop (I have one sitting under the TV with the 3850), you're pretty constrained in upgrades. Very. So, Kaveri fixing the dual GPU crappy show from the Llano era is MAGNIFICENT news for me at least. Specially since GCN should be able to do dual graphics with newer GCN cards (that's the promise, I believe), so there's an "upgrade" path in SFF that I'll take over changing the Video card altogether and try to find a decent one in low profile.

AMD needs to deliver more games that benefit from MANTLE soon. I really really wish AMD pairs up with Codemasters. Man, I'd love the EGO engine to have MANTLE support *_*

Cheers!
 
@yuka: i think the latest catalyst 14.1 beta contains (some) frame pacing fix for older dual gfx combos as well.
http://support.amd.com/en-us/kb-articles/Pages/latest-catalyst-windows-beta.aspx
dx 9 support is still mia. i doubt amd will ever add that despite baiting with "to be added in future".

dual gfx and the apus' biggest advantage has always been in laptops. dual gfx, combined with gpu switching. the other two (intel and nvidia) don't have such capability. apu + dual gfx laptop owners will know better, but from what i've read so far, amd doesn't consistently support laptops with driver updates.
 


i saw that article of tom (when it was published, maybe 1 or 2 year ago) in which they used RAMDisk to utilize os invisible memory on 32bit os and then they create a swap file on that RAMDisk, which is indeed not a wastage or RAM and in no way a bad idea.
bad idea would be to use RAMDisk to store swapfile on 64bit os provided that you don't need extra drive space or using ssd (specially small capacity ssd) or your hdd is too slow (this is my case 😛 ) (espacially when you own top tier version of window 7 that is the "Ultimate" version 😀 in which you can access upto 192GB of RAM)



i didn't read those articles, but if that developer means to say about both 32 and 64bit then :na: he/she needs some researching/studying/testing :kaola: , it is in no way a bad idea in 32bit os :kaola:

what will happen if we create a 2GB pagefile on c drive in 32bit os and use all the invisible memory for another pagefile ? means how memory allocation will occur ? how the system will use multiple pagefiles in case of out of main memory situation ? :heink:
does this multiple page file system will bypass that memalloc loophole of swap on ram method (or whatever it is known as :pt1cable: )



this happens to me when i run 7zip with 6 threads and over 192MB dictionary size 😀 it give me "the system cannot allocate the required amount of memory" error to which my reaction is
9b3.png
 


Just leave it on "system managed", unless you are working with HUGE datasets where you have to manually set the pagefile to some obscene value.

i didn't read those articles, but if that developer means to say about both 32 and 64bit then :na: he/she needs some researching/studying/testing :kaola: , it is in no way a bad idea in 32bit os :kaola:
Click to expand...

Yes it is, since the process of creating the RAMDisk eats up all your free Addresses. Putting the pagefile on a RAMDisk is ALWAYS idiotic.

this happens to me when i run 7zip with 6 threads and over 192MB dictionary size 😀 it give me "the system cannot allocate the required amount of memory" error to which my reaction is
Click to expand...

Which handles fine with a pagefile.
 
^ some softwares like dataram ramdisk or radeon ramdisk use addressable space but some like primo ramdisk uses non addressed ram without using same amount of physical addresses though it needs an image file of that ramdisk on hdd,
and afaik primocache does not need any space on os addressable addresses or any image file on hdd to use os invisible memory for cache (though in 64bit, since os is already capable of addressing all the ram (upto 192GB) thus you won't have any non addressable ram left so you will need to use addressed ram and thats how it meant to be)

thus if you use something like primo ramdisk or cache then you are not stealing system addressed addresses on a 32bit os and you can use page file on this ramdisk which won't slow down your pc (due to slow hdd) when you run in out of memory situation (like multiple 32bit 7zip at a time).
though i don't use ramdisk 😛, since all ramdisk needs image file on hdd thus i don't use it 😛 instead i use primocache as extra read or write cache for my slow hdd to asset my hdd during fraps or gaming sessions 😀

so much addresses in few lines :pt1cable:
 


"Non Addressed RAM". That's a new one.
 


Still my main point which is getting said 16GB DDR3 DIMMs as it does not look like Kaveri supports ECC modules which are the only ones that have 16GB DIMMs right now so max you can do with consumer grade DDR3 non ECC is 32GB.

And until they actually state or put it out, I don't think consumers will see 16GB DDR3 DIMMs. We probably will with DDR4 as I have read that 4GB is the starting DIMM size so it will probably go 4, 8 and 16GB and possibly 32GB, although I doubt in the consumer space we will see 32GB DIMMs.

It's a great feature but useless. It is much like the original Athlon 64s. 64bit was a great feature to brag about except there wasn't a decent 64bit OS to go with it as XP 64 was very new and was not nearly as stable and well rounded as 7 64 is.



The question is what do you do then? Most common users will never see that with 8GB. Hence for the majority, 16GB is overkill.

We enthusiasts however don't do normal things. While most people may listen to music and browse the web, we play games while downloading, watching movies, reading forums and encoding crap all at the same time.



I set mine to 1GB as I never do more than what I have, 16GB. It works out fine for me and save space on my SSD.
 

:chaudar:
ramdisk softwares call it System invisible memory :vip: which is indeed that part of ram which system cannot use because of its address limit in 32bit thus said that it is non addressed :spamafote: :miam:
 


ECC is always more expensive. I remember building a set of four server systems with 128GB DDR3 ECC for a client and the RAM alone was $4000 for all of it. That's about $1000 for 128GB and this is back when 16GB was about $50 for normal which means about $400 for 128GB of normal DDR3.



I saw that today and about lost it. I thought with sales maybe they would list them for near MSRP and I might pick one up but nope. They jumped it up another $200.

AMD can sell Mantle all they want but there is no way justifying $900 bucks for a R9 290X with a good cooler on it.

If they don't find a way to get the pricing in check, like say make crypto mining versions and ones that don't crypto mine for gamers then they will lose a very important market.

I have heard of a site where you can get one for $600 but I have to research if it is legit or not. So far it seems legit but still $900 for a GPU that is beat by one for $750......
 

at least mantle is free to download and use.
with the new retail prices and mining craze, amd has no reason to lower msrp. not only that, amd is bringing back old, phased out gpus to take advantage of the craze. i am guessing anything with 1k shaders and higher and 256bit memory bus will get scooped up by the hopeful miners regardless of price. this has effectively killed the ongoing price war between amd and nvidia. now nvidia doesn't have to lower gfx card prices and amd doesn't have to bundle free games, hype isv relationships and update drivers, let alone lower prices. in the end, no one, incl. buyers outside u.s., will be able to buy higher end amd cards at a cheaper price for the time being. in other countries, price might remain closer to msrp, but they won't drop as before.
 


It is not AMD or say Asus upping the price but rather retailers. Asus stated the MSRP on the R9 290X DC2 is $579.99 and as I said I did find a site that is selling it for $598 which means that e-tailers like Newegg and such are doing the price gouging and effectively screwing their biggest market, enthusiasts.

It is also giving NVidia an advantage because if they keep their prices as is or drop them, more enthusiasts will buy them and when the crypto mining craze dies down, AMD might not be able to recover.

I am upset with Newegg TBH as I have always praised them, even when Amazon Prime had superior service. Newegg used to be where you went to get the best deals, cheaper than stores or other e-tailers.

Now they are just grubbing the money instead of providing a service and AMD is not going to do anything about it meaning if I want a R9 290X, which I do, then I am probably screwed.
 


for mining :clin:
 
New details about Steamroller architecture given by AMD

http://translate.google.com/translate?sl=ja&tl=en&js=n&prev=_t&hl=en&ie=UTF-8&u=http%3A%2F%2Fpc.watch.impress.co.jp%2Fdocs%2Fcolumn%2Fkaigai%2F20140214_635132.html&act=url

http://forwardthinking.pcmag.com/chips/320731-ivytown-steamroller-14-and-16nm-process-highlight-isscc

07.png


8350rocks, check the last part of the second link, because it also discusses tech process:

The Next Generation of Chip Process Tech
There were also a couple of presentations on the next generation of chip process technology, as almost all of the major chip makers have plans to move to 3D or FinFET production, at the 14 or 16nm node (following Intel, which is already shipping 22nm chips with such technology).
Click to expand...
 
Status
Not open for further replies.

TRENDING THREADS

Latest posts

Moderators online

Share this page

COMPANY
RESOURCES
FOLLOW
Top Bottom